jmlr jmlr2012 knowledge-graph by maker-knowledge-mining
1 jmlr-2012-A Case Study on Meta-Generalising: A Gaussian Processes Approach
Author: Grigorios Skolidis, Guido Sanguinetti
Abstract: We propose a novel model for meta-generalisation, that is, performing prediction on novel tasks based on information from multiple different but related tasks. The model is based on two coupled Gaussian processes with structured covariance function; one model performs predictions by learning a constrained covariance function encapsulating the relations between the various training tasks, while the second model determines the similarity of new tasks to previously seen tasks. We demonstrate empirically on several real and synthetic data sets both the strengths of the approach and its limitations due to the distributional assumptions underpinning it. Keywords: transfer learning, meta-generalising, multi-task learning, Gaussian processes, mixture of experts
2 jmlr-2012-A Comparison of the Lasso and Marginal Regression
Author: Christopher R. Genovese, Jiashun Jin, Larry Wasserman, Zhigang Yao
Abstract: The lasso is an important method for sparse, high-dimensional regression problems, with efficient algorithms available, a long history of practical success, and a large body of theoretical results supporting and explaining its performance. But even with the best available algorithms, finding the lasso solutions remains a computationally challenging task in cases where the number of covariates vastly exceeds the number of data points. Marginal regression, where each dependent variable is regressed separately on each covariate, offers a promising alternative in this case because the estimates can be computed roughly two orders faster than the lasso solutions. The question that remains is how the statistical performance of the method compares to that of the lasso in these cases. In this paper, we study the relative statistical performance of the lasso and marginal regression for sparse, high-dimensional regression problems. We consider the problem of learning which coefficients are non-zero. Our main results are as follows: (i) we compare the conditions under which the lasso and marginal regression guarantee exact recovery in the fixed design, noise free case; (ii) we establish conditions under which marginal regression provides exact recovery with high probability in the fixed design, noise free, random coefficients case; and (iii) we derive rates of convergence for both procedures, where performance is measured by the number of coefficients with incorrect sign, and characterize the regions in the parameter space recovery is and is not possible under this metric. In light of the computational advantages of marginal regression in very high dimensional problems, our theoretical and simulations results suggest that the procedure merits further study. Keywords: high-dimensional regression, lasso, phase diagram, regularization
3 jmlr-2012-A Geometric Approach to Sample Compression
Author: Benjamin I.P. Rubinstein, J. Hyam Rubinstein
Abstract: The Sample Compression Conjecture of Littlestone & Warmuth has remained unsolved for a quarter century. While maximum classes (concept classes meeting Sauer’s Lemma with equality) can be compressed, the compression of general concept classes reduces to compressing maximal classes (classes that cannot be expanded without increasing VC dimension). Two promising ways forward are: embedding maximal classes into maximum classes with at most a polynomial increase to VC dimension, and compression via operating on geometric representations. This paper presents positive results on the latter approach and a first negative result on the former, through a systematic investigation of finite maximum classes. Simple arrangements of hyperplanes in hyperbolic space are shown to represent maximum classes, generalizing the corresponding Euclidean result. We show that sweeping a generic hyperplane across such arrangements forms an unlabeled compression scheme of size VC dimension and corresponds to a special case of peeling the one-inclusion graph, resolving a recent conjecture of Kuzmin & Warmuth. A bijection between finite maximum classes and certain arrangements of piecewise-linear (PL) hyperplanes in either a ball or Euclidean space is established. Finally we show that d-maximum classes corresponding to PL-hyperplane arrangements in Rd have cubical complexes homeomorphic to a d-ball, or equivalently complexes that are manifolds with boundary. A main result is that PL arrangements can be swept by a moving hyperplane to unlabeled d-compress any finite maximum class, forming a peeling scheme as conjectured by Kuzmin & Warmuth. A corollary is that some d-maximal classes cannot be embedded into any maximum class of VC-dimension d + k, for any constant k. The construction of the PL sweeping involves Pachner moves on the one-inclusion graph, corresponding to moves of a hyperplane across the intersection of d other hyperplanes. This extends the well known Pachner moves for triangulations to c
4 jmlr-2012-A Kernel Two-Sample Test
Author: Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Schölkopf, Alexander Smola
Abstract: We propose a framework for analyzing and comparing distributions, which we use to construct statistical tests to determine if two samples are drawn from different distributions. Our test statistic is the largest difference in expectations over functions in the unit ball of a reproducing kernel Hilbert space (RKHS), and is called the maximum mean discrepancy (MMD). We present two distributionfree tests based on large deviation bounds for the MMD, and a third test based on the asymptotic distribution of this statistic. The MMD can be computed in quadratic time, although efficient linear time approximations are available. Our statistic is an instance of an integral probability metric, and various classical metrics on distributions are obtained when alternative function classes are used in place of an RKHS. We apply our two-sample tests to a variety of problems, including attribute matching for databases using the Hungarian marriage method, where they perform strongly. Excellent performance is also obtained when comparing distributions over graphs, for which these are the first such tests. ∗. †. ‡. §. Also at Gatsby Computational Neuroscience Unit, CSML, 17 Queen Square, London WC1N 3AR, UK. This work was carried out while K.M.B. was with the Ludwig-Maximilians-Universit¨ t M¨ nchen. a u This work was carried out while M.J.R. was with the Graz University of Technology. Also at The Australian National University, Canberra, ACT 0200, Australia. c 2012 Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Sch¨ lkopf and Alexander Smola. o ¨ G RETTON , B ORGWARDT, R ASCH , S CH OLKOPF AND S MOLA Keywords: kernel methods, two-sample test, uniform convergence bounds, schema matching, integral probability metric, hypothesis testing
Author: Michael W. Mahoney, Lorenzo Orecchia, Nisheeth K. Vishnoi
Abstract: The second eigenvalue of the Laplacian matrix and its associated eigenvector are fundamental features of an undirected graph, and as such they have found widespread use in scientific computing, machine learning, and data analysis. In many applications, however, graphs that arise have several local regions of interest, and the second eigenvector will typically fail to provide information fine-tuned to each local region. In this paper, we introduce a locally-biased analogue of the second eigenvector, and we demonstrate its usefulness at highlighting local properties of data graphs in a semi-supervised manner. To do so, we first view the second eigenvector as the solution to a constrained optimization problem, and we incorporate the local information as an additional constraint; we then characterize the optimal solution to this new problem and show that it can be interpreted as a generalization of a Personalized PageRank vector; and finally, as a consequence, we show that the solution can be computed in nearly-linear time. In addition, we show that this locally-biased vector can be used to compute an approximation to the best partition near an input seed set in a manner analogous to the way in which the second eigenvector of the Laplacian can be used to obtain an approximation to the best partition in the entire input graph. Such a primitive is useful for identifying and refining clusters locally, as it allows us to focus on a local region of interest in a semi-supervised manner. Finally, we provide a detailed empirical evaluation of our method by showing how it can applied to finding locally-biased sparse cuts around an input vertex seed set in social and information networks. Keywords: spectral graph partitioning, local spectral algorithms, Laplacian matrix, semi-supervised learning, personalized pagerank c 2012 Michael W. Mahoney and Lorenzo Orecchia and Nisheeth K. Vishnoi. M AHONEY, O RECCHIA , AND V ISHNOI
Author: Aleix Martinez, Shichuan Du
Abstract: In cognitive science and neuroscience, there have been two leading models describing how humans perceive and classify facial expressions of emotion—the continuous and the categorical model. The continuous model defines each facial expression of emotion as a feature vector in a face space. This model explains, for example, how expressions of emotion can be seen at different intensities. In contrast, the categorical model consists of C classifiers, each tuned to a specific emotion category. This model explains, among other findings, why the images in a morphing sequence between a happy and a surprise face are perceived as either happy or surprise but not something in between. While the continuous model has a more difficult time justifying this latter finding, the categorical model is not as good when it comes to explaining how expressions are recognized at different intensities or modes. Most importantly, both models have problems explaining how one can recognize combinations of emotion categories such as happily surprised versus angrily surprised versus surprise. To resolve these issues, in the past several years, we have worked on a revised model that justifies the results reported in the cognitive science and neuroscience literature. This model consists of C distinct continuous spaces. Multiple (compound) emotion categories can be recognized by linearly combining these C face spaces. The dimensions of these spaces are shown to be mostly configural. According to this model, the major task for the classification of facial expressions of emotion is precise, detailed detection of facial landmarks rather than recognition. We provide an overview of the literature justifying the model, show how the resulting model can be employed to build algorithms for the recognition of facial expression of emotion, and propose research directions in machine learning and computer vision researchers to keep pushing the state of the art in these areas. We also discuss how the model can aid in stu
7 jmlr-2012-A Multi-Stage Framework for Dantzig Selector and LASSO
Author: Ji Liu, Peter Wonka, Jieping Ye
Abstract: We consider the following sparse signal recovery (or feature selection) problem: given a design matrix X ∈ Rn×m (m ≫ n) and a noisy observation vector y ∈ Rn satisfying y = Xβ∗ + ε where ε is the noise vector following a Gaussian distribution N(0, σ2 I), how to recover the signal (or parameter vector) β∗ when the signal is sparse? The Dantzig selector has been proposed for sparse signal recovery with strong theoretical guarantees. In this paper, we propose a multi-stage Dantzig selector method, which iteratively refines the target signal β∗ . We show that if X obeys a certain condition, then with a large probability ˆ the difference between the solution β estimated by the proposed method and the true solution β∗ measured in terms of the ℓ p norm (p ≥ 1) is bounded as ˆ β − β∗ p ≤ C(s − N)1/p log m + ∆ σ, where C is a constant, s is the number of nonzero entries in β∗ , the risk of the oracle estimator ∆ is independent of m and is much smaller than the first term, and N is the number of entries of β∗ larger √ than a certain value in the order of O (σ log m). The proposed √ method improves the estimation √ bound of the standard Dantzig selector approximately from Cs1/p log mσ to C(s − N)1/p log mσ where the value N depends on the number of large entries in β∗ . When N = s, the proposed algorithm achieves the oracle solution with a high probability, where the oracle solution is the projection of the observation vector y onto true features. In addition, with a large probability, the proposed method can select the same number of correct features under a milder condition than the Dantzig selector. Finally, we extend this multi-stage procedure to the LASSO case. Keywords: multi-stage, Dantzig selector, LASSO, sparse signal recovery
8 jmlr-2012-A Primal-Dual Convergence Analysis of Boosting
Author: Matus Telgarsky
Abstract: Boosting combines weak learners into a predictor with low empirical risk. Its dual constructs a high entropy distribution upon which weak learners and training labels are uncorrelated. This manuscript studies this primal-dual relationship under a broad family of losses, including the exponential loss of AdaBoost and the logistic loss, revealing: • Weak learnability aids the whole loss family: for any ε > 0, O (ln(1/ε)) iterations suffice to produce a predictor with empirical risk ε-close to the infimum; • The circumstances granting the existence of an empirical risk minimizer may be characterized in terms of the primal and dual problems, yielding a new proof of the known rate O (ln(1/ε)); • Arbitrary instances may be decomposed into the above two, granting rate O (1/ε), with a matching lower bound provided for the logistic loss. Keywords: boosting, convex analysis, weak learnability, coordinate descent, maximum entropy
9 jmlr-2012-A Topic Modeling Toolbox Using Belief Propagation
Author: Jia Zeng
Abstract: Latent Dirichlet allocation (LDA) is an important hierarchical Bayesian model for probabilistic topic modeling, which attracts worldwide interests and touches on many important applications in text mining, computer vision and computational biology. This paper introduces a topic modeling toolbox (TMBP) based on the belief propagation (BP) algorithms. TMBP toolbox is implemented by MEX C++/Matlab/Octave for either Windows 7 or Linux. Compared with existing topic modeling packages, the novelty of this toolbox lies in the BP algorithms for learning LDA-based topic models. The current version includes BP algorithms for latent Dirichlet allocation (LDA), authortopic models (ATM), relational topic models (RTM), and labeled LDA (LaLDA). This toolbox is an ongoing project and more BP-based algorithms for various topic models will be added in the near future. Interested users may also extend BP algorithms for learning more complicated topic models. The source codes are freely available under the GNU General Public Licence, Version 1.0 at https://mloss.org/software/view/399/. Keywords: topic models, belief propagation, variational Bayes, Gibbs sampling
Author: José Hernández-Orallo, Peter Flach, Cèsar Ferri
Abstract: Many performance metrics have been introduced in the literature for the evaluation of classification performance, each of them with different origins and areas of application. These metrics include accuracy, unweighted accuracy, the area under the ROC curve or the ROC convex hull, the mean absolute error and the Brier score or mean squared error (with its decomposition into refinement and calibration). One way of understanding the relations among these metrics is by means of variable operating conditions (in the form of misclassification costs and/or class distributions). Thus, a metric may correspond to some expected loss over different operating conditions. One dimension for the analysis has been the distribution for this range of operating conditions, leading to some important connections in the area of proper scoring rules. We demonstrate in this paper that there is an equally important dimension which has so far received much less attention in the analysis of performance metrics. This dimension is given by the decision rule, which is typically implemented as a threshold choice method when using scoring models. In this paper, we explore many old and new threshold choice methods: fixed, score-uniform, score-driven, rate-driven and optimal, among others. By calculating the expected loss obtained with these threshold choice methods for a uniform range of operating conditions we give clear interpretations of the 0-1 loss, the absolute error, the Brier score, the AUC and the refinement loss respectively. Our analysis provides a comprehensive view of performance metrics as well as a systematic approach to loss minimisation which can be summarised as follows: given a model, apply the threshold choice methods that correspond with the available information about the operating condition, and compare their expected losses. In order to assist in this procedure we also derive several connections between the aforementioned performance metrics, and we highlight the role of calibra
Author: Neil D. Lawrence
Abstract: We introduce a new perspective on spectral dimensionality reduction which views these methods as Gaussian Markov random fields (GRFs). Our unifying perspective is based on the maximum entropy principle which is in turn inspired by maximum variance unfolding. The resulting model, which we call maximum entropy unfolding (MEU) is a nonlinear generalization of principal component analysis. We relate the model to Laplacian eigenmaps and isomap. We show that parameter fitting in the locally linear embedding (LLE) is approximate maximum likelihood MEU. We introduce a variant of LLE that performs maximum likelihood exactly: Acyclic LLE (ALLE). We show that MEU and ALLE are competitive with the leading spectral approaches on a robot navigation visualization and a human motion capture data set. Finally the maximum likelihood perspective allows us to introduce a new approach to dimensionality reduction based on L1 regularization of the Gaussian random field via the graphical lasso.
12 jmlr-2012-Active Clustering of Biological Sequences
Author: Konstantin Voevodski, Maria-Florina Balcan, Heiko Röglin, Shang-Hua Teng, Yu Xia
Abstract: Given a point set S and an unknown metric d on S, we study the problem of efficiently partitioning S into k clusters while querying few distances between the points. In our model we assume that we have access to one versus all queries that given a point s ∈ S return the distances between s and all other points. We show that given a natural assumption about the structure of the instance, we can efficiently find an accurate clustering using only O(k) distance queries. Our algorithm uses an active selection strategy to choose a small set of points that we call landmarks, and considers only the distances between landmarks and other points to produce a clustering. We use our procedure to cluster proteins by sequence similarity. This setting nicely fits our model because we can use a fast sequence database search program to query a sequence against an entire data set. We conduct an empirical study that shows that even though we query a small fraction of the distances between the points, we produce clusterings that are close to a desired clustering given by manual classification. Keywords: clustering, active clustering, k-median, approximation algorithms, approximation stability, clustering accuracy, protein sequences ∗. A preliminary version of this article appeared under the title Efficient Clustering with Limited Distance Information in the Proceedings of the Twenty-Sixth Conference on Uncertainty in Artificial Intelligence, AUAI Press, Corvallis, Oregon, 632-641. †. Most of this work was completed at Boston University. c 2012 Konstantin Voevodski, Maria-Florina Balcan, Heiko R¨ glin, Shang-Hua Teng and Yu Xia. o ¨ VOEVODSKI , BALCAN , R OGLIN , T ENG AND X IA
13 jmlr-2012-Active Learning via Perfect Selective Classification
Author: Ran El-Yaniv, Yair Wiener
Abstract: We discover a strong relation between two known learning models: stream-based active learning and perfect selective classification (an extreme case of ‘classification with a reject option’). For these models, restricted to the realizable case, we show a reduction of active learning to selective classification that preserves fast rates. Applying this reduction to recent results for selective classification, we derive exponential target-independent label complexity speedup for actively learning general (non-homogeneous) linear classifiers when the data distribution is an arbitrary high dimensional mixture of Gaussians. Finally, we study the relation between the proposed technique and existing label complexity measures, including teaching dimension and disagreement coefficient. Keywords: classification with a reject option, perfect classification, selective classification, active learning, selective sampling, disagreement coefficient, teaching dimension, exploration vs. exploitation 1. Introduction and Related Work Active learning is an intriguing learning model that provides the learning algorithm with some control over the learning process, potentially leading to significantly faster learning. In recent years it has been gaining considerable recognition as a vital technique for efficiently implementing inductive learning in many industrial applications where abundance of unlabeled data exists, and/or in cases where labeling costs are high. In this paper we expose a strong relation between active learning and selective classification, another known alternative learning model (Chow, 1970; El-Yaniv and Wiener, 2010). Focusing on binary classification in realizable settings we consider standard stream-based active learning, which is also referred to as online selective sampling (Atlas et al., 1990; Cohn et al., 1994). In this model the learner is given an error objective ε and then sequentially receives unlabeled examples. At each step, after observing an unlabeled example x, the learner decides whether or not to request the label of x. The learner should terminate the learning process and output a binary classifier whose true error is guaranteed to be at most ε with high probability. The penalty incurred by the learner is the number of label requests made and this number is called the label complexity. A label complexity bound of O(d log(d/ε)) for actively learning ε-good classifier from a concept class with VC-dimension d, provides an exponential speedup in terms of 1/ε relative to standard (passive) supervised learning where the sample complexity is typically O(d/ε). The study of (stream-based, realizable) active learning is paved with very interesting theoretical results. Initially, only a few cases were known where active learning provides significant advanc 2012 Ran El-Yaniv and Yair Wiener. E L -YANIV AND W IENER tage over passive learning. Perhaps the most favorable result was an exponential label complexity speedup for learning homogeneous linear classifiers where the (linearly separable) data is uniformly distributed over the unit sphere. This result was manifested by various authors using various analysis techniques, for a number of strategies that can all be viewed in hindsight as approximations or variations of the “CAL algorithm” of Cohn et al. (1994). Among these studies, the earlier theoretical results (Seung et al., 1992; Freund et al., 1993, 1997; Fine et al., 2002; Gilad-Bachrach, 2007) considered Bayesian settings and studied the speedup obtained by the Query by Committee (QBC) algorithm. The more recent results provided PAC style analyses (Dasgupta et al., 2009; Hanneke, 2007a, 2009). Lack of positive results for other non-toy problems, as well as various additional negative results that were discovered, led some researchers to believe that active learning is not necessarily advantageous in general. Among the striking negative results is Dasgupta’s negative example for actively learning general (non-homogeneous) linear classifiers (even in two dimensions) under the uniform distribution over the sphere (Dasgupta, 2005). A number of recent innovative papers proposed alternative models for active learning. Balcan et al. (2008) introduced a subtle modification of the traditional label complexity definition, which opened up avenues for new positive results. According to their new definition of “non-verifiable” label complexity, the active learner is not required to know when to stop the learning process with a guaranteed ε-good classifier. Their main result, under this definition, is that active learning is asymptotically better than passive learning in the sense that only o(1/ε) labels are required for actively learning an ε-good classifier from a concept class that has a finite VC-dimension. Another result they accomplished is an exponential label complexity speedup for (non-verifiable) active learning of non-homogeneous linear classifiers under the uniform distribution over the the unit sphere. Based on Hanneke’s characterization of active learning in terms of the “disagreement coefficient” (Hanneke, 2007a), Friedman (2009) recently extended the Balcan et al. results and proved that a target-dependent exponential speedup can be asymptotically achieved for a wide range of “smooth” learning problems (in particular, the hypothesis class, the instance space and the distribution should all be expressible by smooth functions). He proved that under such smoothness conditions, for any target hypothesis h∗ , Hanneke’s disagreement coefficient is bounded above in terms of a constant c(h∗ ) that depends on the unknown target hypothesis h∗ (and is independent of δ and ε). The resulting label complexity is O (c(h∗ ) d polylog(d/ε)) (Hanneke, 2011b). This is a very general result but the target-dependent constant involved in this bound is only guaranteed to be finite. With this impressive progress in the case of target-dependent bounds for active learning, the current state of affairs in the target-independent bounds for active learning arena leaves much to be desired. To date the most advanced result in this model, which was already essentially established by Seung et al. and Freund et al. more than fifteen years ago (Seung et al., 1992; Freund et al., 1993, 1997), is still a target-independent exponential speed up bound for homogeneous linear classifiers under the uniform distribution over the sphere. The other learning model we contemplate that will be shown to have strong ties to active learning, is selective classification, which is mainly known in the literature as ‘classification with a reject option.’ This old-timer model, that was already introduced more than fifty years ago (Chow, 1957, 1970), extends standard supervised learning by allowing the classifier to opt out from predictions in cases where it is not confident. The incentive is to increase classification reliability over instances that are not rejected by the classifier. Thus, using selective classification one can potentially achieve 256 ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION a lower error rate using the same labeling “budget.” The main quantities that characterize a selective classifier are its (true) error and coverage rate (or its complement, the rejection rate). There is already substantial volume of research publications on selective classification, that kept emerging through the years. The main theme in many of these publications is the implementation of certain reject mechanisms for specific learning algorithms like support vector machines and neural networks. Among the few theoretical studies on selective classification, there are various excess risk bounds for ERM learning (Herbei and Wegkamp, 2006; Bartlett and Wegkamp, 2008; Wegkamp, 2007), and certain coverage/risk guarantees for selective ensemble methods (Freund et al., 2004). In a recent work (El-Yaniv and Wiener, 2010) the trade-off between error and coverage was examined and in particular, a new extreme case of selective learning was introduced. In this extreme case, termed here “perfect selective classification,” the classifier is given m labeled examples and is required to instantly output a classifier whose true error is perfectly zero with certainty. This is of course potentially doable only if the classifier rejects a sufficient portion of the instance space. A non-trivial result for perfect selective classification is a high probability lower bound on the classifier coverage (or equivalently, an upper bound on its rejection rate). Such bounds have recently been presented in El-Yaniv and Wiener (2010). In Section 3 we present a reduction of active learning to perfect selective classification that preserves “fast rates.” This reduction enables the luxury of analyzing dynamic active learning problems as static problems. Relying on a recent result on perfect selective classification from El-Yaniv and Wiener (2010), in Section 4 we then apply our reduction and conclude that general (non-homogeneous) linear classifiers are actively learnable at exponential (in 1/ε) label complexity rate when the data distribution is an arbitrary unknown finite mixture of high dimensional Gaussians. While we obtain exponential label complexity speedup in 1/ε, we incur exponential slowdown in d 2 , where d is the problem dimension. Nevertheless, in Section 5 we prove a lower bound of Ω((log m)(d−1)/2 (1 + o(1)) on the label complexity, when considering the class of unrestricted linear classifiers under a Gaussian distribution. Thus, an exponential slowdown in d is unavoidable in such settings. Finally, in Section 6 we relate the proposed technique to other complexity measures for active learning. Proving and using a relation to the teaching dimension (Goldman and Kearns, 1995) we show, by relying on a known bound for the teaching dimension, that perfect selective classification with meaningful coverage can be achieved for the case of axis-aligned rectangles under a product distribution. We then focus on Hanneke’s disagreement coefficient and show that the coverage of perfect selective classification can be bounded below using the disagreement coefficient. Conversely, we show that the disagreement coefficient can be bounded above using any coverage bound for perfect selective classification. Consequently, the results here imply that the disagreement coefficient can be sufficiently bounded to ensure fast active learning for the case of linear classifiers under a mixture of Gaussians. 2. Active Learning and Perfect Selective Classification In binary classification the goal is to learn an accurate binary classifier, h : X → {±1}, from a finite labeled training sample. Here X is some instance space and the standard assumption is that the training sample, Sm = {(xi , yi )}m , containing m labeled examples, is drawn i.i.d. from some i=1 unknown distribution P(X,Y ) defined over X × {±1}. The classifier h is chosen from some hypothesis class H . In this paper we focus on the realizable setting whereby labels are defined by 257 E L -YANIV AND W IENER some unknown target hypothesis h∗ ∈ H . Thus, the underlying distribution reduces to P(X). The performance of a classifier h is quantified by its true zero-one error, R(h) Pr{h(X) = h∗ (X)}. A positive result for a classification problem (H , P) is a learning algorithm that given an error target ε and a confidence parameter δ can output, based on Sm , an hypothesis h whose error R(h) ≤ ε, with probability of at least 1 − δ. A bound B(ε, δ) on the size m of labeled training sample sufficient for achieving this is called the sample complexity of the learning algorithm. A classical result is that any consistent learning algorithm has sample complexity of O( 1 (d log( 1 ) + log( 1 ))), where d is ε ε δ the VC-dimension of H (see, e.g., Anthony and Bartlett, 1999). 2.1 Active Learning We consider the following standard active learning model. In this model the learner sequentially observes unlabeled instances, x1 , x2 , . . ., that are sampled i.i.d. from P(X). After receiving each xi , the learning algorithm decides whether or not to request its label h∗ (xi ), where h∗ ∈ H is an unknown target hypothesis. Before the start of the game the algorithm is provided with some desired error rate ε and confidence level δ. We say that the learning algorithm actively learned the problem instance (H , P) if at some point it can terminate this process, after observing m instances and requesting k labels, and output an hypothesis h ∈ H whose error R(h) ≤ ε, with probability of at least 1 − δ. The quality of the algorithm is quantified by the number k of requested labels, which is called the label complexity. A positive result for a learning problem (H , P) is a learning algorithm that can actively learn this problem for any given ε and δ, and for every h∗ , with label complexity bounded above by L(ε, δ, h∗ ). If there is a label complexity bound that is O(polylog(1/ε)) we say that the problem is actively learnable at exponential rate. 2.2 Selective Classification Following the formulation in El-Yaniv and Wiener (2010) the goal in selective classification is to learn a pair of functions (h, g) from a labeled training sample Sm (as defined above for passive learning). The pair (h, g), which is called a selective classifier, consists of a binary classifier h ∈ H , and a selection function, g : X → {0, 1}, which qualifies the classifier h as follows. For any sample x ∈ X , the output of the selective classifier is (h, g)(x) h(x) iff g(x) = 1, and (h, g)(x) abstain iff g(x) = 0. Thus, the function g is a filter that determines a sub-domain of X over which the selective classifier will abstain from classifications. A selective classifier is thus characterized by its coverage, Φ(h, g) EP {g(x)}, which is the P-weighted volume of the sub-domain of X that is not filtered out, and its error, R(h, g) = E{I(h(X) = h∗ (X)) · g(X)}/Φ(h, g), which is the zero-one loss restricted to the covered sub-domain. Note that this is a “smooth” generalization of passive learning and, in particular, R(h, g) reduces to R(h) (standard classification) if g(x) ≡ 1. We expect to see a trade-off between R(h, g) and Φ(h, g) in the sense that smaller error should be obtained by compromising the coverage. A major issue in selective classification is how to optimally control this trade-off. In this paper we are concerned with an extreme case of this trade-off whereby (h, g) is required to achieve a perfect score of zero error with certainty. This extreme learning objective is termed perfect learning in El-Yaniv and Wiener (2010). Thus, for a perfect selective classifier (h, g) we always have R(h, g) = 0, and its quality is determined by its guaranteed coverage. A positive result for (perfect) selective classification problem (H , P) is a learning algorithm that uses a labeled training sample Sm (as in passive learning) to output a perfect selective classifier (h, g) for which Φ(h, g) ≥ BΦ (H , δ, m) with probability of at least 1 − δ, for any given δ. The bound 258 ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION BΦ = BΦ (H , δ, m) is called a coverage bound (or coverage rate) and its complement, 1 − BΦ , is called a rejection bound (or rate). A coverage rate BΦ = 1 − O( polylog(m) ) (and the corresponding m 1 − BΦ rejection rate) are qualified as fast. 2.3 The CAL Algorithm and the Consistent Selective Strategy (CSS) The major players in active learning and in perfect selective classification are the CAL algorithm and the consistent selective strategy (CSS), respectively. To define them we need the following definitions. Definition 1 (Version space, Mitchell, 1977) Given an hypothesis class H and a training sample Sm , the version space V SH ,Sm is the set of all hypotheses in H that classify Sm correctly. Definition 2 (Disagreement set, Hanneke, 2007a; El-Yaniv and Wiener, 2010) Let G ⊂ H . The disagreement set w.r.t. G is defined as DIS(G ) {x ∈ X : ∃h1 , h2 ∈ G The agreement set w.r.t. G is AGR(G ) s.t. h1 (x) = h2 (x)} . X \ DIS(G ). The main strategy for active learning in the realizable setting (Cohn et al., 1994) is to request labels only for instances belonging to the disagreement set and output any (consistent) hypothesis belonging to the version space. This strategy is often called the CAL algorithm. A related strategy for perfect selective classification was proposed in El-Yaniv and Wiener (2010) and termed consistent selective strategy (CSS). Given a training set Sm , CSS takes the classifier h to be any hypothesis in V SH ,Sm (i.e., a consistent learner), and takes a selection function g that equals one for all points in the agreement set with respect to V SH ,Sm , and zero otherwise. 3. From Coverage Bound to Label Complexity Bound In this section we present a reduction from stream-based active learning to perfect selective classification. Particularly, we show that if there exists for H a perfect selective classifier with a fast rejection rate of O(polylog(m)/m), then the CAL algorithm will actively learn H with exponential label complexity rate of O(polylog(1/ε)). Lemma 3 Let Sm = {(x1 , y1 ), . . . , (xm , ym )} be a sequence of m labeled samples drawn i.i.d. from an unknown distribution P(X) and let Si = {(x1 , y1 ), . . . , (xi , yi )} be the i-prefix of Sm . Then, with probability of at least 1 − δ over random choices of Sm , the following bound holds simultaneously for all i = 1, . . . , m − 1, Pr xi+1 ∈ DIS(V SH ,Si )|Si ≤ 1 − BΦ H , δ , 2⌊log2 (i)⌋ , log2 (m) where BΦ (H , δ, m) is a coverage bound for perfect selective classification with respect to hypothesis class H , confidence δ and sample size m . 259 E L -YANIV AND W IENER Proof For j = 1, . . . , m, abbreviate DIS j DIS(V SH ,S j ) and AGR j AGR(V SH ,S j ). By definition, DIS j = X \ AGR j . By the definitions of a coverage bound and agreement/disagreement sets, with probability of at least 1 − δ over random choices of S j BΦ (H , δ, j) ≤ Pr{x ∈ AGR j |S j } = Pr{x ∈ DIS j |S j } = 1 − Pr{x ∈ DIS j |S j }. Applying the union bound we conclude that the following inequality holds simultaneously with high probability for t = 0, . . . , ⌊log2 (m)⌋ − 1, Pr{x2t +1 ∈ DIS2t |S2t } ≤ 1 − BΦ H , δ , 2t . log2 (m) (1) For all j ≤ i, S j ⊆ Si , so DISi ⊆ DIS j . Therefore, since the samples in Sm are all drawn i.i.d., for any j ≤ i, Pr {xi+1 ∈ DISi |Si } ≤ Pr xi+1 ∈ DIS j |S j = Pr x j+1 ∈ DIS j |S j . The proof is complete by setting j = 2⌊log2 (i)⌋ ≤ i, and applying inequality (1). Lemma 4 (Bernstein’s inequality Hoeffding, 1963) Let X1 , . . . , Xn be independent zero-mean random variables. Suppose that |Xi | ≤ M almost surely, for all i. Then, for all positive t, n 2 /2 t . Pr ∑ Xi > t ≤ exp − 2 + Mt/3 i=1 ∑E Xj Lemma 5 Let Zi , i = 1, . . . , m, be independent Bernoulli random variables with success probabilities pi . Then, for any 0 < δ < 1, with probability of at least 1 − δ, m ∑ (Zi − E{Zi }) ≤ 2 ln i=1 Proof Define Wi 1 2 1 ∑ pi + 3 ln δ . δ Zi − E{Zi } = Zi − pi . Clearly, E{Wi } = 0, |Wi | ≤ 1, E{Wi2 } = pi (1 − pi ). Applying Bernstein’s inequality (Lemma 4) on the Wi , n t 2 /2 t 2 /2 = exp − Pr ∑ Wi > t ≤ exp − ∑ pi (1 − pi ) + t/3 i=1 ∑ E W j2 + t/3 ≤ exp − t 2 /2 . ∑ pi + t/3 Equating the right-hand side to δ and solving for t, we have t 2 /2 1 = ln δ ∑ pi + t/3 ⇐⇒ 2 1 1 t 2 − t · ln − 2 ln ∑ pi = 0, 3 δ δ 260 ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION and the positive solution of this quadratic equation is t= 1 1 ln + 3 δ 1 21 1 2 1 ln + 2 ln ∑ pi < ln + 9 δ δ 3 δ 2 ln 1 pi . δ∑ Lemma 6 Let Z1 , Z2 , . . . , Zm be a high order Markov sequence of dependent binary random variables defined in the same probability space. Let X1 , X2 , . . . , Xm be a sequence of independent random variables such that, Pr {Zi = 1|Zi−1 , . . . , Z1 , Xi−1 , . . . , X1 } = Pr {Zi = 1|Xi−1 , . . . , X1 } . Define P1 Pr {Z1 = 1}, and for i = 2, . . . , m, Pi Pr {Zi = 1|Xi−1 , . . . , X1 } . Let b1 , b2 . . . bm be given constants independent of X1 , X2 , . . . , Xm .1 Assume that Pi ≤ bi simultaneously for all i with probability of at least 1 − δ/2, δ ∈ (0, 1). Then, with probability of at least 1 − δ, m m 2 2 2 ∑ Zi ≤ ∑ bi + 2 ln δ ∑ bi + 3 ln δ . i=1 i=1 We proceed with a direct proof of Lemma 6. An alternative proof of this lemma, using supermartingales, appears in Appendix B. Proof For i = 1, . . . , m, let Wi be binary random variables satisfying bi + I(Pi ≤ bi ) · (Pi − bi ) , Pi bi − Pi ,0 , Pr{Wi = 1|Zi = 0, Xi−1 , . . . , X1 } max 1 − Pi Pr{Wi = 1|Wi−1 , . . . ,W1 , Xi−1 , . . . , X1 } = Pr{Wi = 1|Xi−1 , . . . , X1 }. Pr{Wi = 1|Zi = 1, Xi−1 , . . . , X1 } We notice that Pr{Wi = 1|Xi−1 , . . . , X1 } = Pr{Wi = 1, Zi = 1|Xi−1 , . . . , X1 } + Pr{Wi = 1, Zi = 0|Xi−1 , . . . , X1 } = Pr{Wi = 1|Zi = 1, Xi−1 , . . . , X1 } Pr{Zi = 1|Xi−1 , . . . , X1 } + Pr{Wi = 1|Zi = 0, Xi−1 , . . . , X1 } Pr{Zi = 0|Xi−1 , . . . , X1 } = Pi + bi −Pii (1 − Pi ) = bi , Pi ≤ bi ; 1−P bi · Pi + 0 = bi , else. Pi Hence the distribution of each Wi is independent of Xi−1 , . . . , X1 , and the Wi are independent Bernoulli random variables with success probabilities bi . By construction if Pi ≤ bi then Pr{Wi = 1|Zi = 1} = X Pr{Wi = 1|Zi = 1, Xi−1 , . . . , X1 } = 1. 1. Precisely we require that each of the bi were selected before Xi are chosen 261 E L -YANIV AND W IENER By assumption Pi ≤ bi for all i simultaneously with probability of at least 1−δ/2. Therefore, Zi ≤ Wi simultaneously with probability of at least 1 − δ/2. We now apply Lemma 5 on the Wi . The proof is then completed using the union bound. Theorem 7 Let Sm be a sequence of m unlabeled samples drawn i.i.d. from an unknown distribution P. Then with probability of at least 1 − δ over choices of Sm , the number of label requests k by the CAL algorithm is bounded by k ≤ Ψ(H , δ, m) + where Ψ(H , δ, m) 2 2 2 2 ln Ψ(H , δ, m) + ln , δ 3 δ m ∑ i=1 1 − BΦ H , δ , 2⌊log2 (i)⌋ 2 log2 (m) and BΦ (H , δ, m) is a coverage bound for perfect selective classification with respect to hypothesis class H , confidence δ and sample size m . Proof According to CAL, the label of sample xi will be requested iff xi ∈ DIS(V SH ,Si−1 ). For i = 1, . . . , m, let Zi be binary random variables such that Zi 1 iff CAL requests a label for sample xi . Applying Lemma 3 we get that for all i = 2, . . . , m, with probability of at least 1 − δ/2 Pr{Zi = 1|Si−1 } = Pr xi ∈ DIS(V SH ,Si−1 )|Si−1 ≤ 1 − BΦ H , δ , 2⌊log2 (i−1)⌋ . 2 log2 (m) For i = 1, BΦ (H , δ, 1) = 0 and the above inequality trivially holds. An application of Lemma 6 on the variables Zi completes the proof. Theorem 7 states an upper bound on the label complexity expressed in terms of m, the size of the sample provided to CAL. This upper bound is very convenient for directly analyzing the active learning speedup relative to supervised learning. A standard label complexity upper bound, which depends on 1/ε, can be extracted using the following simple observation. Lemma 8 (Hanneke, 2009; Anthony and Bartlett, 1999) Let Sm be a sequence of m unlabeled samples drawn i.i.d. from an unknown distribution P. Let H be a hypothesis class whose finite VC dimension is d, and let ε and δ be given. If m≥ 4 2 12 d ln + ln , ε ε δ then, with probability of at least 1 − δ, CAL will output a classifier whose true error is at most ε. Proof Hanneke (2009) observed that since CAL requests a label whenever there is a disagreement in the version space, it is guaranteed that after processing m examples, CAL will output a classifier that is consistent with all the m examples introduced to it. Therefore, CAL is a consistent learner. A classical result (Anthony and Bartlett, 1999, Theorem 4.8) is that any consistent learner will achieve, with probability of at least 1 − δ, a true error not exceeding ε after observing at most 12 2 4 ε d ln ε + ln δ labeled examples. 262 ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION Theorem 9 Let H be a hypothesis class whose finite VC dimension is d. If the rejection rate of CSS polylog( m ) δ (see definition in Section 2.3) is O , then (H , P) is actively learnable with exponential m label complexity speedup. Proof Plugging this rejection rate into Ψ (defined in Theorem 7) we have, m m polylog δ Ψ(H , δ, m) ∑ 1 − BΦ (H , , 2⌊log2 (i)⌋ ) = ∑ O log2 (m) i i=1 i=1 Applying Lemma 41 we get Ψ(H , δ, m) = O polylog By Theorem 7, k = O polylog m δ m log(m) δ i log(m) δ . . , and an application of Lemma 8 concludes the proof. 4. Label Complexity Bounding Technique and Its Applications In this section we present a novel technique for deriving target-independent label complexity bounds for active learning. The technique combines the reduction of Theorem 7 and a general datadependent coverage bound for selective classification from El-Yaniv and Wiener (2010). For some learning problems it is a straightforward technical exercise, involving VC-dimension calculations, to arrive with exponential label complexity bounds. We show a few applications of this technique resulting in both reproductions of known label complexity exponential rates as well as a new one. The following definitions (El-Yaniv and Wiener, 2010) are required for introducing the technique. Definition 10 (Version space compression set) For any hypothesis class H , let Sm be a labeled sample of m points inducing a version space V SH ,Sm . The version space compression set, S′ ⊆ Sm , ˆ ˆ is a smallest subset of Sm satisfying V SH ,Sm = V SH ,S′ . The (unique) number n = n(H , Sm ) = |S′ | is called the version space compression set size. Remark 11 Our ”version space compression set” is precisely Hanneke’s ”minimum specifying set” (Hanneke, 2007b) for f on U with respect to V , where, f = h∗ , U = Sm , V = H [Sm ] (see Definition 23). Definition 12 (Characterizing hypothesis) For any subset of hypotheses G ⊆ H , the characterizing hypothesis of G , denoted fG (x), is a binary hypothesis over X (not restricted to H ) obtaining positive values over the agreement set AGR(G ) (Definition 2), and zero otherwise. Definition 13 (Order-n characterizing set) For each n, let Σn be the set of all possible labeled samples of size n (all n-subsets, each with all 2n possible labelings). The order-n characterizing set of H , denoted Fn , is the set of all characterizing hypotheses fG (x), where G ⊆ H is a version space induced by some member of Σn . 263 E L -YANIV AND W IENER Definition 14 (Characterizing set complexity) Let Fn be the order-n characterizing set of H . The order-n characterizing set complexity of H , denoted γ (H , n), is the VC-dimension of Fn . The following theorem, credited to (El-Yaniv and Wiener, 2010, Theorem 21), is a powerful data-dependent coverage bound for perfect selective learning, expressed in terms of the version space compression set size and the characterizing set complexity. Theorem 15 (Data-dependent coverage guarantee) For any m, let a1 , a2 , . . . , am ∈ R be given, such that ai ≥ 0 and ∑m ai ≤ 1. Let (h, g) be perfect selective classifier (CSS, see Section 2.3). i=1 Then, R(h, g) = 0, and for any 0 ≤ δ ≤ 1, with probability of at least 1 − δ, Φ(h, g) ≥ 1 − 2 γ (H , n) ln+ ˆ m 2em 2 , + ln an δ γ (H , n) ˆ ˆ where n is the size of the version space compression set and γ (H , n) is the order-n characterizing ˆ ˆ ˆ set complexity of H . Given an hypothesis class H , our recipe to deriving active learning label complexity bounds for H is: (i) calculate both n and γ (H , n); (ii) apply Theorem 15, obtaining a bound BΦ for the ˆ ˆ coverage; (iii) plug BΦ in Theorem 7 to get a label complexity bound expressed as a summation; (iv) Apply Lemma 41 to obtain a label complexity bound in a closed form. 4.1 Examples In the following example we derive a label complexity bound for the concept class of thresholds (linear separators in R). Although this is a toy example (for which an exponential rate is well known) it does exemplify the technique, and in many other cases the application of the technique is not much harder. Let H be the class of thresholds. We first show that the corresponding version space compression set size n ≤ 2. Assume w.l.o.g. that h∗ (x) I(x > w) for some w ∈ (0, 1). Let ˆ x− max{xi ∈ Sm |yi = −1} and x+ min(xi ∈ Sm |yi = +1). At least one of x− or x+ exist. Let ′ ′ Sm = {(x− , −1), (x+ , +1)}. Then V SH ,Sm = V SH ,Sm , and n = |Sm | ≤ 2. Now, γ (H , 2) = 2, because ˆ ′ the order-2 characterizing set of H is the class of intervals in R whose VC-dimension is 2. Plugging these numbers in Theorem 15, and using the assignment a1 = a2 = 1/2, BΦ (H , δ, m) = 1 − 2 4 ln (m/δ) 2 ln (em) + ln = 1−O . m δ m Next we plug BΦ in Theorem 7 obtaining a raw label complexity m Ψ(H , δ, m) = ∑ 1 − BΦ H , i=1 δ , 2⌊log2 (i)⌋ 2 log2 (m) m = ∑O i=1 ln (log2 (m) · i/δ) . i Finally, by applying Lemma 41, with a = 1 and b = log2 m/δ, we conclude that Ψ(H , δ, m) = O ln2 m δ . Thus, H is actively learnable with exponential speedup, and this result applies to any distribution. In Table 1 we summarize the n and γ (H , n) values we calculated for four other hypothesis classes. The ˆ ˆ 264 ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION Hypothesis class Distribution n ˆ γ (H , n) ˆ Linear separators in R Intervals in R Linear separators in R2 any any (target-dependent)2 any distribution on the unit circle (target-dependent)2 2 4 4 2 4 4 Linear separators in Rd Balanced axis-aligned rectangles in Rd mixture of Gaussians product distribution O (log m)d−1 /δ O (log (dm/δ)) O nd/2+1 ˆ O (d n log n) ˆ ˆ Table 1: The n and γ of various hypothesis spaces achieving exponential rates. ˆ last two cases are fully analyzed in Sections 4.2 and 6.1, respectively. For the other classes, where γ and n are constants, it is clear (Theorem 15) that exponential rates are obtained. We emphasize that ˆ the bounds for these two classes are target-dependent as they require that Sm include at least one sample from each class. 4.2 Linear Separators in Rd Under Mixture of Gaussians In this section we state and prove our main example, an exponential label complexity bound for linear classifiers in Rd . Theorem 16 Let H be the class of all linear binary classifiers in Rd , and let the underlying distribution be any mixture of a fixed number of Gaussians in Rd . Then, with probability of at least 1 − δ over choices of Sm , the number of label requests k by CAL is bounded by 2 k=O (log m)d +1 δ(d+3)/2 . Therefore by Lemma 8 we get k = O (poly(1/δ) · polylog(1/ε)) . Proof The following is a coverage bound for linear classifiers in d dimensions that holds in our setting with probability of at least 1 − δ (El-Yaniv and Wiener, 2010, Corollary 33),3 2 Φ(h, g) ≥ 1 − O 1 (log m)d · (d+3)/2 m δ . (2) 2. Target-dependent with at least one sample in each class. 3. This bound uses the fact that for linear classifiers in d dimensions n = O (log m)d−1 /δ (El-Yaniv and Wiener, 2010, ˆ Lemma 32), and that γ (H , n) = O nd/2+1 (El-Yaniv and Wiener, 2010, Lemma 27). ˆ ˆ 265 E L -YANIV AND W IENER Plugging this bound in Theorem 7 we obtain, Ψ(H , δ, m) = m ∑ i=1 1 − BΦ H , m = ∑O i=1 = O δ , 2⌊log2 (i)⌋ 2 log2 (m) 2 log2 (m) (log i)d · i δ log2 (m) δ d+3 2 d+3 2 m (log(i))d ·∑ i i=1 2 . Finally, an application of Lemma 41 with a = d 2 and b = 1 completes the proof. 5. Lower Bound on Label Complexity In the previous section we have derived an upper bound on the label complexity of CAL for various classifiers and distributions. In the case of linear classifiers in Rd we have shown an exponential speed up in terms of 1/ε but also an exponential slow down in terms of the dimension d. In passive learning there is a linear dependency in the dimension while in our case (active learning using CAL) there is an exponential one. Is it an artifact of our bounding technique or a fundamental phenomenon? To answer this question we derive an asymptotic lower bound on the label complexity. We show that the exponential dependency in d is unavoidable (at least asymptotically) for every bounding technique when considering linear classifier even under a single Gaussian (isotropic) distribution. The argument is obtained by the observation that CAL has to request a label to any point on the convex hull of a sample Sm . The bound is obtained using known results from probabilistic geometry, which bound the first two moments of the number of vertices of a random polytope under the Gaussian distribution. Definition 17 (Gaussian polytope) Let X1 , ..., Xm be i.i.d. random points in Rd with common stan1 dard normal distribution (with zero mean and covariance matrix 2 Id ). A Gaussian polytope Pm is the convex hull of these random points. Denote by fk (Pm ) the number of k-faces in the Gaussian polytope Pm . Note that f0 (Pm ) is the number of vertices in Pm . The following two Theorems asymptotically bound the average and variance of fk (Pm ). Theorem 18 (Hug et al., 2004, Theorem 1.1) Let X1 , ..., Xm be i.i.d. random points in Rd with common standard normal distribution. Then E fk (Pm ) = c(k,d) (log m) d−1 2 · (1 + o(1)) as m → ∞, where c(k,d) is a constant depending only on k and d. 266 ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION Theorem 19 (Hug and Reitzner, 2005, Theorem 1.1) Let X1 , ..., Xm be i.i.d. random points in Rd with common standard normal distribution. Then there exists a positive constant cd , depending only on the dimension, such that d−1 Var ( fk (Pm )) ≤ cd (log m) 2 for all k ∈ {0, . . . , d − 1}. We can now use Chebyshev’s inequality to lower bound the number of vertices in Pm ( f0 (Pm )) with high probability. Theorem 20 Let X1 , ..., Xm be i.i.d. random points in Rd with common standard normal distribution and δ > 0 be given. Then with probability of at least 1 − δ, f0 (Pm ) ≥ cd (log m) d−1 2 d−1 cd ˜ − √ (log m) 4 δ · (1 + o(1)) as m → ∞, where cd and cd are constants depending only on d. ˜ Proof Using Chebyshev’s inequality (in the second inequality), as well as Theorem 19 we get Pr ( f0 (Pm ) > E f0 (Pm ) − t) = 1 − Pr ( f0 (Pm ) ≤ E f0 (Pm ) − t) ≥ 1 − Pr (| f0 (Pm ) − E f0 (Pm )| ≥ t) d−1 cd Var ( f0 (Pm )) ≥ 1 − 2 (log m) 2 . ≥ 1− 2 t t Equating the RHS to 1 − δ and solving for t we get t= (log m) cd δ d−1 2 . Applying Theorem 18 completes the proof. Theorem 21 (Lower bound) Let H be the class of linear binary classifiers in Rd , and let the underlying distribution be standard normal distribution in Rd . Then there exists a target hypothesis such that, with probability of at least 1 − δ over choices of Sm , the number of label requests k by CAL is bounded by d−1 cd k ≥ (log m) 2 · (1 + o(1)). 2 as m → ∞, where cd is a constant depending only on d. Proof Let us look at the Gaussian polytope Pm induced by the random sample Sm . As long as all labels requested by CAL have the same value (the case of minuscule minority class) we note that every vertex of Pm falls in the region of disagreement with respect to any subset of Sm that do not include that specific vertex. Therefore, CAL will request label at least for each vertex of Pm . For sufficiently large m, in particular, 4 2cd d−1 ˜ √ log m ≥ , cd δ we conclude the proof by applying Theorem 20. 267 E L -YANIV AND W IENER 6. Relation to Existing Label Complexity Measures A number of complexity measures to quantify the speedup in active learning have been proposed. In this section we show interesting relations between our techniques and two well known measures, namely the teaching dimension (Goldman and Kearns, 1995) and the disagreement coefficient (Hanneke, 2009). Considering first the teaching dimension, we prove in Lemma 26 that the version space compression set size is bounded above, with high probability, by the extended teaching dimension growth function (introduced by Hanneke, 2007b). Consequently, it follows that perfect selective classification with meaningful coverage can be achieved for the case of axis-aligned rectangles under a product distribution. We then focus on Hanneke’s disagreement coefficient and show in Theorem 34 that the coverage of CSS can be bounded below using the disagreement coefficient. Conversely, in Corollary 39 we show that the disagreement coefficient can be bounded above using any coverage bound for CSS. Consequently, the results here imply that the disagreement coefficient, θ(ε) grows slowly with 1/ε for the case of linear classifiers under a mixture of Gaussians. 6.1 Teaching Dimension The teaching dimension is a label complexity measure proposed by Goldman and Kearns (1995). The dimension of the hypothesis class H is the minimum number of examples required to present to any consistent learner in order to uniquely identify any hypothesis in the class. We now define the following variation of the extended teaching dimension (Heged¨ s, 1995) u due to Hanneke. Throughout we use the notation h1 (S) = h2 (S) to denote the fact that the two hypotheses agree on the classification of all instances in S. ¨ Definition 22 (Extended Teaching Dimension, Hegedus, 1995; Hanneke, 2007b) Let V ⊆ H , m ≥ m, 0, U ∈ X ∀f ∈ H , XT D( f ,V,U) = inf {t | ∃R ⊆ U : | {h ∈ V : h(R) = f (R)} | ≤ 1 ∧ |R| ≤ t} . Definition 23 (Hanneke, 2007b) For V ⊆ H , V [Sm ] denotes any subset of V such that ∀h ∈ V, | h′ ∈ V [Sm ] : h′ (Sm ) = h(Sm ) | = 1. Claim 24 Let Sm be a sample of size m, H an hypothesis class, and n = n(H , Sm ), the version space ˆ compression set size. Then, XT D(h∗ , H [Sm ], Sm ) = n. ˆ Proof Let Sn ⊆ Sm be a version space compression set. Assume, by contradiction, that there exist ˆ two hypotheses h1 , h2 ∈ H [Sm ], each of which agrees on the given classifications of all examples in Sn . Therefore, h1 , h2 ∈ V SH ,Sn , and by the definition of version space compression set, we get ˆ ˆ h1 , h2 ∈ V SH ,Sm . Hence, | h ∈ H [Sm ] : h(Sm ) = h∗ (Sm ) | ≥ 2, which contradicts definition 23. Therefore, | h ∈ H [Sm ] : h(Sn ) = h∗ (Sn ) | ≤ 1, ˆ ˆ 268 ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION and XT D(h∗ , H [Sm ], Sm ) ≤ |Sn | = n. ˆ ˆ Let R ⊂ Sm be any subset of size |R| < n. Consequently, V SH ,Sm ⊂ V SH ,R , and there exist hypothesis, ˆ ′ ∈ VS h H ,R , that agrees with all labeled examples in R, but disagrees with at least one example in Sm . Thus, h′ (Sm ) = h∗ (Sm ), and according to definition 23, there exist hypotheses h1 , h2 ∈ H [Sm ] such that h1 (Sm ) = h′ (Sm ) = h∗ (Sm ) = h2 (Sm ). But h1 (R) = h2 (R) = h∗ (R), so | {h ∈ V [Sm ] : h(R) = h∗ (R)} | ≥ 2. It follows that XT D(h∗ , H [Sm ], Sm ) ≥ n. ˆ Definition 25 (XTD Growth Function, Hanneke, 2007b) For m ≥ 0, V ⊆ H , δ ∈ [0, 1], XT D(V, P, m, δ) = inf t|∀h ∈ H , Pr {XT D(h,V [Sm ], Sm ) > t} ≤ δ . Lemma 26 Let H be an hypothesis class, P an unknown distribution, and δ > 0. Then, with probability of at least 1 − δ, n ≤ XT D(H , P, m, δ). ˆ Proof According to Definition 25, with probability of at least 1 − δ, XT D(h∗ , H [Sm ], Sm ) ≤ XT D(H , P, m, δ). Applying Claim 24 completes the proof. Lemma 27 (Balanced Axis-Aligned Rectangles, Hanneke, 2007b, Lemma 4) If P is a product distribution on Rd with continuous CDF, and H is the set of axis-aligned rectangles such that ∀h ∈ H , PrX∼P {h(X) = +1} ≥ λ, then, XT D(H , P, m, δ) ≤ O d2 dm log . λ δ Lemma 28 Blumer et al., 1989, Lemma 3.2.3 Let F be a binary hypothesis class of finite VC dimension d ≥ 1. For all k ≥ 1, define the k-fold union, Fk∪ Then, for all k ≥ 1, ∪k f i : f i ∈ F , 1 ≤ i ≤ k . i=1 VC(Fk∪ ) ≤ 2dk log2 (3k). Lemma 29 (order-n characterizing set complexity) Let H be the class of axis-aligned rectangles in Rd . Then, γ(H , n) ≤ O (dn log n) . 269 E L -YANIV AND W IENER − + Proof Let Sn = Sk ∪ Sn−k be a sample of size n composed of k negative examples, {x1 , x2 , . . . xk }, and n − k positive ones. Let H be the class of axis-aligned rectangles. We define, ∀1 ≤ i ≤ k, + Sn−k ∪ {(xi , −1)} . Ri Notice that V SH ,Ri includes all axis aligned rectangles that classify all samples in S+ as positive, and xi as negative. Therefore, the agreement region of V SH ,Ri is composed of two components as depicted in Figure 1. The first component is the smallest rectangle that bounds the positive samples, and the second is an unbounded convex polytope defined by up to d hyperplanes intersecting at xi . Let AGRi be the agreement region of V SH ,Ri and AGR the agreement region of V SH ,Sn . Clearly, Ri ⊆ Sn , so V SH ,Sn ⊆ V SH ,Ri , and AGRi ⊆ AGR, and it follows that k i=1 AGRi ⊆ AGR. Assume, by contradiction, that x ∈ AGR but x ∈ k AGRi . Therefore, for any 1 ≤ i ≤ k, there exist i=1 (i) (i) (i) (i) two hypotheses h1 , h2 ∈ V SH ,Ri , such that, h1 (x) = h2 (x). Assume, without loss of generality, (i) that h1 (x) = 1. We define k h1 (i) h1 k and (i) h2 , h2 i=1 i=1 (i) meaning that h1 classifies a sample as positive if and only if all hypotheses h1 classify it as positive. Noting that the intersection of axis-aligned rectangles is itself an axis-aligned rectangle, we know (i) (i) that h1 , h2 ∈ H . Moreover, for any xi we have, h1 (xi ) = h2 (xi ) = −1, so also h1 (xi ) = h2 (xi ) = −1, and h1 , h2 ∈ V SH ,Sn . But h1 (x) = h2 (x). Contradiction. Therefore, k AGRi . AGR = i=1 It is well known that the VC dimension of a hyper-rectangle in Rd is 2d. The VC dimension of AGRi is bounded by the VC dimension of the union of two hyper-rectangles in Rd . Furthermore, the VC dimension of AGR is bounded by the VC dimension of the union of all AGRi . Applying Lemma 28 twice we get, VCdim {AGR} ≤ 42dk log2 (3k) ≤ 42dn log2 (3n). If k = 0 then the entire sample is positive and the region of agreement is an hyper-rectangle. Therefore, VCdim {AGR} = 2d. If k = n then the entire sample is negative and the region of agreement is the points of the samples themselves. Hence, VCdim {AGR} = n. Overall we get that in all cases, VCdim {AGR} ≤ 42dn log2 (3n) = O (dn log n) . 270 ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION Figure 1: Agreement region of V SH ,Ri . Corollary 30 (Balanced Axis-Aligned Rectangles) Under the same conditions of Lemma 27, the class of balanced axis-aligned rectangles in Rd can be perfectly selectively learned with fast coverage rate. Proof Applying Lemmas 26 and 27 we get that with probability of at least 1 − δ, dm d2 log . λ δ n≤O ˆ Any balanced axis-aligned rectangle belongs to the class of all axis-aligned rectangles. Therefore, the coverage of CSS for the class of balanced axis-aligned rectangles is bounded bellow by the coverage of the class of axis-aligned rectangles. Applying Lemma 29, and assuming m ≥ d, we obtain, γ (H , n) ≤ O d ˆ d2 d2 dm dm log log log λ δ λ δ ≤O dm d3 log2 . λ λδ Applying Theorem 15 completes the proof. 6.2 Disagreement Coefficient In this section we show interesting relations between the disagreement coefficient and coverage bounds in perfect selective classification. We begin by defining, for an hypothesis h ∈ H , the set of all hypotheses that are r-close to h. Definition 31 (Hanneke, 2011b, p.337) For any hypothesis h ∈ H , distribution P over X , and r > 0, define the set B(h, r) of all hypotheses that reside in a ball of radius r around h, B(h, r) h′ ∈ H : Pr X∼P h′ (X) = h(X) ≤ r . Theorem 32 (Vapnik and Chervonenkis, 1971; Anthony and Bartlett, 1999, p.53) Let H be a hypothesis class with VC-dimension d. For any probability distribution P on X × {±1}, with probability of at least 1 − δ over the choice of Sm , any hypothesis h ∈ H consistent with Sm satisfies R(h) ≤ η(d, m, δ) 2 2em 2 + ln . d ln m d δ 271 E L -YANIV AND W IENER For any G ⊆ H and distribution P we denote by ∆G the volume of the disagreement region of G, ∆G Pr {DIS(G)} . Definition 33 (Disagreement coefficient, Hanneke, 2009) Let ε ≥ 0. The disagreement coefficient of the hypothesis class H with respect to the target distribution P is θ(ε) θh∗ (ε) = sup r>ε ∆B(h∗ , r) . r The following theorem formulates an intimate relation between active learning (disagreement coefficient) and selective classification. Theorem 34 Let H be an hypothesis class with VC-dimension d, P an unknown distribution, ε ≥ 0, and θ(ε), the corresponding disagreement coefficient. Let (h, g) be a perfect selective classifier (CSS, see Section 2.3). Then, R(h, g) = 0, and for any 0 ≤ δ ≤ 1, with probability of at least 1 − δ, Φ(h, g) ≥ 1 − θ(ε) · max {η(d, m, δ), ε} . Proof Clearly, R(h, g) = 0, and it remains to prove the coverage bound. By Theorem 32, with probability of at least 1 − δ, ∀h ∈ V SH ,Sm R(h) ≤ η(d, m, δ) ≤ max {η(d, m, δ), ε} . Therefore, V SH ,Sm ⊆ B (h∗ , max {η(d, m, δ), ε}) , ∆V SH ,Sm ≤ ∆B (h∗ , max {η(d, m, δ), ε}) . By Definition 33, for any r′ > ε, ∆B(h∗ , r′ ) ≤ θ(ε)r′ . Thus, the proof is complete by recalling that Φ(h, g) = 1 − ∆V SH ,Sm . Theorem 34 tells us that whenever our learning problem (specified by the pair (H , P)) has a disagreement coefficient that grows slowly with respect to 1/ε , it can be (perfectly) selectively learned with a “fast” coverage bound. Consequently, through Theorem 9 we also know that in each case where there exists a disagreement coefficient that grows slowly with respect to 1/ε, active learning with a fast rate can also be deduced directly through a reduction from perfect selective classification. It follows that as far as fast rates in active learning are concerned, whatever can be accomplished by bounding the disagreement coefficient, can be accomplished also using perfect selective classification. This result is summarized in the following corollary. Corollary 35 Let H be an hypothesis class with VC-dimension d, P an unknown distribution, and θ(ε), the corresponding disagreement coefficient. If θ(ε) = O(polylog(1/ε)), there exists a coverage bound such that an application of Theorem 7 ensures that (H , P) is actively learnable with exponential label complexity speedup. 272 ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION Proof The proof is established by straightforward applications of Theorems 34 with ε = 1/m and 9. The following result, due to Hanneke (2011a), implies a coverage upper bound for CSS. Lemma 36 (Hanneke, 2011a, Proof of Lemma 47) Let H be an hypothesis class, P an unknown distribution, and r ∈ (0, 1). Then, EP ∆Dm ≥ (1 − r)m ∆B (h∗ , r) , where Dm V SH ,Sm ∩ B (h∗ , r) . (3) Theorem 37 (Coverage upper bound) Let H be an hypothesis class, P an unknown distribution, and δ ∈ (0, 1). Then, for any r ∈ (0, 1), 1 > α > δ, BΦ (H , δ, m) ≤ 1 − where BΦ (H , δ, m) is any coverage bound. (1 − r)m − α ∆B (h∗ , r) , 1−α Proof Recalling the definition of Dm (3), clearly Dm ⊆ V SH ,Sm and Dm ⊆ B(h∗ , r). These inclusions imply (respectively), by the definition of disagreement set, ∆Dm ≤ ∆V SH ,Sm , and ∆Dm ≤ ∆B(h∗ , r). (4) Using Markov’s inequality (in inequality (5) of the following derivation) and applying (4) (in equality (6)), we thus have, (1 − r)m − α (1 − r)m − α ∆B (h∗ , r) ≤ Pr ∆Dm ≤ ∆B (h∗ , r) 1−α 1−α 1 − (1 − r)m Pr ∆B (h∗ , r) − ∆Dm ≥ ∆B (h∗ , r) 1−α 1 − (1 − r)m Pr |∆B (h∗ , r) − ∆Dm | ≥ ∆B (h∗ , r) 1−α E {|∆B (h∗ , r) − ∆Dm |} (1 − α) · (1 − (1 − r)m ) ∆B (h∗ , r) ∆B (h∗ , r) − E∆Dm . (1 − α) · (1 − (1 − r)m ) ∆B (h∗ , r) Pr ∆V SH ,Sm ≤ = ≤ ≤ = Applying Lemma 36 we therefore obtain, ≤ (1 − α) · ∆B (h∗ , r) − (1 − r)m ∆B(h∗ , r) = 1 − α < 1 − δ. (1 − (1 − r)m ) ∆B (h∗ , r) Observing that for any coverage bound, Pr ∆V SH ,Sm ≤ 1 − BΦ (H , δ, m) ≥ 1 − δ, completes the proof. 273 (5) (6) E L -YANIV AND W IENER Corollary 38 Let H be an hypothesis class, P an unknown distribution, and δ ∈ (0, 1/8). Then for any m ≥ 2, 1 1 , BΦ (H , δ, m) ≤ 1 − ∆B h∗ , 7 m where BΦ (H , δ, m) is any coverage bound. Proof The proof is established by a straightforward application of Theorem 37 with α = 1/8 and r = 1/m. With Corollary 38 we can bound the disagreement coefficient for settings whose coverage bound is known. Corollary 39 Let H be an hypothesis class, P an unknown distribution, and BΦ (H , δ, m) a coverage bound. Then the disagreement coefficient is bounded by, θ(ε) ≤ max sup 7 · r∈(ε,1/2) 1 − BΦ (H , 1/9, ⌊1/r⌋) ,2 . r Proof Applying Corollary 38 we get that for any r ∈ (0, 1/2), 1 − BΦ (H , 1/9, ⌊1/r⌋) ∆B(h∗ , r) ∆B(h∗ , 1/⌊1/r⌋) ≤ ≤ 7· . r r r Therefore, θ(ε) = sup r>ε ∆B(h∗ , r) ≤ max r sup 7 · r∈(ε,1/2) 1 − BΦ (H , 1/9, ⌊1/r⌋) ,2 . r Corollary 40 Let H be the class of all linear binary classifiers in Rd , and let the underlying distribution be any mixture of a fixed number of Gaussians in Rd . Then θ(ε) ≤ O polylog 1 ε . Proof Applying Corollary 39 together with inequality 2 we get that θ(ε) ≤ max sup 7 · r∈(ε,1/2) 1 − BΦ (H , 1/9, ⌊1/r⌋) ,2 r 7 ≤ max sup ·O r∈(ε,1/2) r 2 d+3 (log ⌊1/r⌋)d ·9 2 ⌊1/r⌋ 274 ,2 ≤O 1 log ε d2 . ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION 7. Concluding Remarks For quite a few years, since its inception, the theory of target-independent bounds for noise-free active learning managed to handle relatively simple settings, mostly revolving around homogeneous linear classifiers under the uniform distribution over the sphere. It is likely that this distributional uniformity assumption was often adapted to simplify analyses. However, it was shown by Dasgupta (2005) that under this distribution, exponential speed up cannot be achieved when considering general (non homogeneous) linear classifiers. The reason for this behavior is related to the two tasks that a good active learner should successfully accomplish: exploration and exploitation. Intuitively (and oversimplifying things) exploration is the task of obtaining at least one sample in each class, and exploitation is the process of refining the decision boundary by requesting labels of points around the boundary. Dasgupta showed that exploration cannot be achieved fast enough under the uniform distribution on the sphere. The source of this difficulty is the fact that under this distribution all training points reside on their convex hull. In general, the speed of exploration (using linear classifiers) depends on the size (number of vertices) of the convex hull of the training set. When using homogeneous linear classifiers, exploration is trivially achieved (under the uniform distribution) and exploitation can achieve exponential speedup. So why in the non-verifiable model (Balcan et al., 2008) it is possible to achieve exponential speedup even when using non homogeneous linear classifiers under the uniform distribution? The answer is that in the non-verifiable model, label complexity attributed to exploration is encapsulated in a target-dependent “constant.” Specifically, in Balcan et al. (2008) this constant is explicitly defined to be the probability mass of the minority class. Indeed, in certain noise free settings using linear classifiers, where the minority class is large enough, exploration is a non issue. In general, however, exploration is a major bottleneck in practical active learning (Baram et al., 2004; Begleiter et al., 2008). The present results show how exponential speedup can be achieved, including exploration, when using different (and perhaps more natural) distributions. With these good news, a somewhat pessimistic picture arises from the lower bound we obtained for the exponential dependency on the dimension d. This negative result is not restricted to streambased active learning and readily applies also to the pool-based model. While the bound is only asymptotic, we conjecture that it also holds for finite samples. Moreover, we believe that within the stream- or pool-based settings a similar statement should hold true for any active learning method (and not necessarily CAL-based querying strategies). This result indicates that when performing noise free active learning of linear classifiers, aggressive feature selection is beneficial for exploration speedup. We note, however, that it remains open whether a slowdown exponent of d (rather than d 2 ) is achievable. We have exposed interesting relations of the present technique to well known complexity measures for active learning, namely, the teaching dimension and the disagreement coefficient. These developments were facilitated by observations made by Hanneke on the teaching dimension and the disagreement coefficient. These relations gave rise to further observations on active learning, which are discussed in Section 6 and include exponential speedup for balanced axis-aligned rectangles. Finally, we note that the intimate relation between selective classification and the disagreement coefficient was recently exposed in another result for selective classification where the disagreement coefficient emerged as a dominating factor in a coverage bound for agnostic selective classification (El-Yaniv and Wiener, 2011). 275 E L -YANIV AND W IENER Acknowledgments We thank the anonymous reviewers for their good comments. This paper particularly benefited from insightful observations made by one of the reviewers, which are summarized in Section 6, including the proof of Theorem 37 and the link between our n and the extended teaching dimension ˆ (Lemmas 26 and 27). Appendix A. Lemma 41 For any m ≥ 3, a ≥ 1, b ≥ 1 we get lna (bi) i m ∑ i=1 Proof Setting f (x) lna (bx) x , < 4 a+1 ln (b(m + 1)). a we have lna−1 (bx) df = (a − ln bx) · . dx x2 Therefore, f is monotonically increasing when x < ea /b, monotonically decreasing function when x ≥ ea /b and its attains its maximum at x = ea /b. Consequently, for i < ea /b − 1, or i ≥ ea /b + 1, i+1 f (i) ≤ f (x)dx. x=i−1 For ea /b − 1 ≤ i < ea /b + 1, f (i) ≤ f (ea /b) = b a e a ≤ aa . (7) Therefore, if m < ea − 1 we have, m ∑ i=1 m f (i) = lna (b) + ∑ f (i) < 2 · i=2 m+1 x=1 f (x)dx ≤ 2 lna+1 (b(m + 1)). a+1 Otherwise, m ≥ ea /b, in which case we overcome the change of slope by adding twice the (upper bound on the) maximal value (7), m ∑ f (i) < i=1 ≤ 2 2 2 lna+1 (b(m + 1)) + 2aa = lna+1 (b(m + 1)) + aa+1 a+1 a+1 a 2 4 2 lna+1 (b(m + 1)) + lna+1 bm ≤ lna+1 (b(m + 1)). a+1 a a 276 ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION Appendix B. Alternative Proof of Lemma 6 Using Super Martingales Define Wk ∑k (Zi − bi ). We assume that with probability of at least 1 − δ/2, i=1 Pr{Zi |Z1 , . . . , Zi−1 } ≤ bi , simultaneously for all i. Since Zi is a binary random variable it is easy to see that (w.h.p.), EZi {Wi |Z1 , . . . , Zi−1 } = Pr{Zi |Z1 , . . . , Zi−1 } − bi +Wi−1 ≤ Wi−1 , and the sequence W1m W1 , . . . ,Wm is a super-martingale with high probability. We apply the following theorem by McDiarmid that refers to martingales (but can be shown to apply to supermartingales, by following its original proof). Theorem 42 (McDiarmid, 1998, Theorem 3.12) Let Y1 , . . . ,Yn be a martingale difference sequence with −ak ≤ Yk ≤ 1 − ak for each k; let A = 1 ∑ ak . Then, for any ε > 0, n Pr ∑ Yk ≥ Anε ≤ exp (−[(1 + ε) ln(1 + ε) − ε]An) ≤ exp − Anε2 . 2(1 + ε/3) In our case, Yk = Wk −Wk−1 = Zk − bk ≤ 1 − bk and we apply the (revised) theorem with ak and An ∑ bk B. We thus obtain, for any 0 < ε < 1, Pr ∑ Zk ≥ B + Bε ≤ exp − bk Bε2 . 2(1 + ε/3) Equating the right-hand side to δ/2, we obtain ε = 2 2 ln ± 3 δ 4 22 2 ln + 8B ln 9 δ δ ≤ 1 2 ln + 3 δ 1 22 ln + 9 δ = 2 2 ln + 3 δ 2B ln 2 δ 2B ln /2B 2 δ /B /B. Applying the union bound completes the proof. References M. Anthony and P.L. Bartlett. Neural Network Learning; Theoretical Foundations. Cambridge University Press, 1999. L. Atlas, D. Cohn, R. Ladner, A.M. El-Sharkawi, and R.J. Marks. Training connectionist networks with queries and selective sampling. In Neural Information Processing Systems (NIPS), pages 566–573, 1990. M.F. Balcan, S. Hanneke, and J. Wortman. The true sample complexity of active learning. In 21st Annual Conference on Learning Theory (COLT), pages 45–56, 2008. 277 E L -YANIV AND W IENER Y. Baram, R. El-Yaniv, and K. Luz. Online choice of active learning algorithms. Journal of Machine Learning Research, 5:255–291, 2004. P.L. Bartlett and M.H. Wegkamp. Classification with a reject option using a hinge loss. Journal of Machine Learning Research, 9:1823–1840, 2008. R. Begleiter, R. El-Yaniv, and D. Pechyony. Repairing self-confident active-transductive learners using systematic exploration. Pattern Recognition Letters, 29(9):1245–1251, 2008. A. Blumer, A. Ehrenfeucht, D. Haussler, and M.K. Warmuth. Chervonenkis dimension. Journal of the ACM, 36, 1989. Learnability and the Vapnik- C.K. Chow. An optimum character recognition system using decision function. IEEE Transactions on Computers, 6(4):247–254, 1957. C.K. Chow. On optimum recognition error and reject trade-off. IEEE Transactions on Information Theory, 16:41–36, 1970. D. Cohn, L. Atlas, and R. Ladner. Improving generalization with active learning. Machine Learning, 15(2):201–221, 1994. S. Dasgupta. Coarse sample complexity bounds for active learning. In Advances in Neural Information Processing Systems 18, pages 235–242, 2005. S. Dasgupta, A. Tauman Kalai, and C. Monteleoni. Analysis of perceptron-based active learning. Journal of Machine Learning Research, 10:281–299, 2009. R. El-Yaniv and Y. Wiener. On the foundations of noise-free selective classification. Journal of Machine Learning Research, 11:1605–1641, 2010. R. El-Yaniv and Y. Wiener. Agnostic selective classification. In Neural Information Processing Systems (NIPS), 2011. S. Fine, R. Gilad-Bachrach, and E. Shamir. Query by committee, linear separation and random walks. Theoretical Computer Science, 284(1):25–51, 2002. Y. Freund, H.S. Seung, E. Shamir, and N. Tishby. Information, prediction, and Query by Committee. In Advances in Neural Information Processing Systems (NIPS) 5, pages 483–490, 1993. Y. Freund, H.S. Seung, E. Shamir, and N. Tishby. Selective sampling using the query by committee algorithm. Machine Learning, 28:133–168, 1997. Y. Freund, Y. Mansour, and R.E. Schapire. Generalization bounds for averaged classifiers. Annals of Statistics, 32(4):1698–1722, 2004. E. Friedman. Active learning for smooth problems. In Proceedings of the 22nd Annual Conference on Learning Theory (COLT), 2009. R. Gilad-Bachrach. To PAC and Beyond. PhD thesis, the Hebrew University of Jerusalem, 2007. S. Goldman and M. Kearns. On the complexity of teaching. JCSS: Journal of Computer and System Sciences, 50, 1995. 278 ACTIVE L EARNING VIA P ERFECT S ELECTIVE C LASSIFICATION S. Hanneke. A bound on the label complexity of agnostic active learning. In ICML ’07: Proceedings of the 24th international conference on Machine learning, pages 353–360, 2007a. S. Hanneke. Teaching dimension and the complexity of active learning. In Proceedings of the 20th Annual Conference on Learning Theory (COLT), volume 4539 of Lecture Notes in Artificial Intelligence, pages 66–81, 2007b. S. Hanneke. Theoretical Foundations of Active Learning. PhD thesis, Carnegie Mellon University, 2009. S. Hanneke. Activized learning: Transforming passive to active with improved label complexity. CoRR, abs/1108.1766, 2011a. URL http://arxiv.org/abs/1108.1766. informal publication. S. Hanneke. Rates of convergence in active learning. Annals of Statistics, 37(1):333–361, 2011b. T. Heged¨ s. Generalized teaching dimensions and the query complexity of learning. In COLT: u Proceedings of the Workshop on Computational Learning Theory, Morgan Kaufmann Publishers, 1995. R. Herbei and M.H. Wegkamp. Classification with reject option. The Canadian Journal of Statistics, 34(4):709–721, 2006. W. Hoeffding. Probability inequalities for sums of bounded random variables. Journal of the American Statistical Association, 58(301):13–30, March 1963. D. Hug and M. Reitzner. Gaussian polytopes: variances and limit theorems, June 2005. D. Hug, G. O. Munsonious, and M. Reitzner. Asymptotic mean values of Gaussian polytopes. Beitr¨ ge Algebra Geom., 45:531–548, 2004. a C. McDiarmid. Concentration. In M. Habib, C. McDiarmid, J. Ramirez-Alfonsin, and B. Reed, editors, Probabilistic Methods for Algorithmic Discrete Mathematics, volume 16, pages 195– 248. Springer-Verlag, 1998. T. Mitchell. Version spaces: a candidate elimination approach to rule learning. In IJCAI’77: Proceedings of the 5th international joint conference on Artificial Intelligence, pages 305–310, 1977. H.S. Seung, M. Opper, and H. Sompolinsky. Query by committee. In Proceedings of the Fifth Annual Workshop on Computational Learning theory (COLT), pages 287–294, 1992. V. Vapnik and A. Chervonenkis. On the uniform convergence of relative frequencies of events to their probabilities. Theory of Probability and its Applications, 16:264–280, 1971. M.H. Wegkamp. Lasso type classifiers with a reject option. Electronic Journal of Statistics, 1: 155–168, 2007. 279
14 jmlr-2012-Activized Learning: Transforming Passive to Active with Improved Label Complexity
Author: Steve Hanneke
Abstract: We study the theoretical advantages of active learning over passive learning. Specifically, we prove that, in noise-free classifier learning for VC classes, any passive learning algorithm can be transformed into an active learning algorithm with asymptotically strictly superior label complexity for all nontrivial target functions and distributions. We further provide a general characterization of the magnitudes of these improvements in terms of a novel generalization of the disagreement coefficient. We also extend these results to active learning in the presence of label noise, and find that even under broad classes of noise distributions, we can typically guarantee strict improvements over the known results for passive learning. Keywords: active learning, selective sampling, sequential design, statistical learning theory, PAC learning, sample complexity 1. Introduction and Background The recent rapid growth in data sources has spawned an equally rapid expansion in the number of potential applications of machine learning methodologies to extract useful concepts from these data. However, in many cases, the bottleneck in the application process is the need to obtain accurate annotation of the raw data according to the target concept to be learned. For instance, in webpage classification, it is straightforward to rapidly collect a large number of webpages, but training an accurate classifier typically requires a human expert to examine and label a number of these webpages, which may require significant time and effort. For this reason, it is natural to look for ways to reduce the total number of labeled examples required to train an accurate classifier. In the traditional machine learning protocol, here referred to as passive learning, the examples labeled by the expert are sampled independently at random, and the emphasis is on designing learning algorithms that make the most effective use of the number of these labeled examples available. However, it is possible to go beyond such methods by altering the protocol itself, allowing the learning algorithm to sequentially select the examples to be labeled, based on its observations of the labels of previously-selected examples; this interactive protocol is referred to as active learning. The objective in designing this selection mechanism is to focus the expert’s efforts toward labeling only the most informative data for the learning process, thus eliminating some degree of redundancy in the information content of the labeled examples. ∗. Some of these (and related) results previously appeared in the author’s doctoral dissertation (Hanneke, 2009b). c 2012 Steve Hanneke. H ANNEKE It is now well-established that active learning can sometimes provide significant practical and theoretical advantages over passive learning, in terms of the number of labels required to obtain a given accuracy. However, our current understanding of active learning in general is still quite limited in several respects. First, since we are lacking a complete understanding of the potential capabilities of active learning, we are not yet sure to what standards we should aspire for active learning algorithms to meet, and in particular this challenges our ability to characterize how a “good” active learning algorithm should behave. Second, since we have yet to identify a complete set of general principles for the design of effective active learning algorithms, in many cases the most effective known active learning algorithms have problem-specific designs (e.g., designed specifically for linear separators, or decision trees, etc., under specific assumptions on the data distribution), and it is not clear what components of their design can be abstracted and transferred to the design of active learning algorithms for different learning problems (e.g., with different types of classifiers, or different data distributions). Finally, we have yet to fully understand the scope of the relative benefits of active learning over passive learning, and in particular the conditions under which such improvements are achievable, as well as a general characterization of the potential magnitudes of these improvements. In the present work, we take steps toward closing this gap in our understanding of the capabilities, general principles, and advantages of active learning. Additionally, this work has a second theme, motivated by practical concerns. To date, the machine learning community has invested decades of research into constructing solid, reliable, and well-behaved passive learning algorithms, and into understanding their theoretical properties. We might hope that an equivalent amount of effort is not required in order to discover and understand effective active learning algorithms. In particular, rather than starting from scratch in the design and analysis of active learning algorithms, it seems desirable to leverage this vast knowledge of passive learning, to whatever extent possible. For instance, it may be possible to design active learning algorithms that inherit certain desirable behaviors or properties of a given passive learning algorithm. In this way, we can use a given passive learning algorithm as a reference point, and the objective is to design an active learning algorithm with performance guarantees strictly superior to those of the passive algorithm. Thus, if the passive learning algorithm has proven effective in a variety of common learning problems, then the active learning algorithm should be even better for those same learning problems. This approach also has the advantage of immediately supplying us with a collection of theoretical guarantees on the performance of the active learning algorithm: namely, improved forms of all known guarantees on the performance of the given passive learning algorithm. Due to its obvious practical advantages, this general line of informal thinking dominates the existing literature on empirically-tested heuristic approaches to active learning, as most of the published heuristic active learning algorithms make use of a passive learning algorithm as a subroutine (e.g., SVM, logistic regression, k-NN, etc.), constructing sets of labeled examples and feeding them into the passive learning algorithm at various times during the execution of the active learning algorithm (see the references in Section 7). Below, we take a more rigorous look at this general strategy. We develop a reduction-style framework for studying this approach to the design of active learning algorithms relative to a given passive learning algorithm. We then proceed to develop and analyze a variety of such methods, to realize this approach in a very general sense. Specifically, we explore the following fundamental questions. 1470 ACTIVIZED L EARNING • Is there a general procedure that, given any passive learning algorithm, transforms it into an active learning algorithm requiring significantly fewer labels to achieve a given accuracy? • If so, how large is the reduction in the number of labels required by the resulting active learning algorithm, compared to the number of labels required by the original passive algorithm? • What are sufficient conditions for an exponential reduction in the number of labels required? • To what extent can these methods be made robust to imperfect or noisy labels? In the process of exploring these questions, we find that for many interesting learning problems, the techniques in the existing literature are not capable of realizing the full potential of active learning. Thus, exploring this topic in generality requires us to develop novel insights and entirely new techniques for the design of active learning algorithms. We also develop corresponding natural complexity quantities to characterize the performance of such algorithms. Several of the results we establish here are more general than any related results in the existing literature, and in many cases the algorithms we develop use significantly fewer labels than any previously published methods. 1.1 Background The term active learning refers to a family of supervised learning protocols, characterized by the ability of the learning algorithm to pose queries to a teacher, who has access to the target concept to be learned. In practice, the teacher and queries may take a variety of forms: a human expert, in which case the queries may be questions or annotation tasks; nature, in which case the queries may be scientific experiments; a computer simulation, in which case the queries may be particular parameter values or initial conditions for the simulator; or a host of other possibilities. In our present context, we will specifically discuss a protocol known as pool-based active learning, a type of sequential design based on a collection of unlabeled examples; this seems to be the most common form of active learning in practical use today (e.g., Settles, 2010; Baldridge and Palmer, 2009; Gangadharaiah, Brown, and Carbonell, 2009; Hoi, Jin, Zhu, and Lyu, 2006; Luo, Kramer, Goldgof, Hall, Samson, Remsen, and Hopkins, 2005; Roy and McCallum, 2001; Tong and Koller, 2001; McCallum and Nigam, 1998). We will not discuss alternative models of active learning, such as online (Dekel, Gentile, and Sridharan, 2010) or exact (Heged¨ s, 1995). In the pool-based active learning u setting, the learning algorithm is supplied with a large collection of unlabeled examples (the pool), and is allowed to select any example from the pool to request that it be labeled. After observing the label of this example, the algorithm can then select another unlabeled example from the pool to request that it be labeled. This continues sequentially for a number of rounds until some halting condition is satisfied, at which time the algorithm returns a function intended to approximately mimic and generalize the observed labeling behavior. This setting contrasts with passive learning, where the learning algorithm is supplied with a collection of labeled examples without any interaction. Supposing the labels received agree with some true target concept, the objective is to use this returned function to approximate the true target concept on future (previously unobserved) data points. The hope is that, by carefully selecting which examples should be labeled, the algorithm can achieve improved accuracy while using fewer labels compared to passive learning. The motivation for this setting is simple. For many modern machine learning problems, unlabeled examples are inexpensive and available in abundance, while annotation is time-consuming or expensive. For instance, this is the case in the aforementioned webpage classification problem, where the pool would 1471 H ANNEKE be the set of all webpages, and labeling a webpage requires a human expert to examine the website content. Settles (2010) surveys a variety of other applications for which active learning is presently being used. To simplify the discussion, in this work we focus specifically on binary classification, in which there are only two possible labels. The results generalize naturally to multiclass classification as well. As the above description indicates, when studying the advantages of active learning, we are primarily interested in the number of label requests sufficient to achieve a given accuracy, a quantity referred to as the label complexity (Definition 1 below). Although active learning has been an active topic in the machine learning literature for many years now, our theoretical understanding of this topic was largely lacking until very recently. However, within the past few years, there has been an explosion of progress. These advances can be grouped into two categories: namely, the realizable case and the agnostic case. 1.1.1 T HE R EALIZABLE C ASE In the realizable case, we are interested in a particularly strict scenario, where the true label of any example is determined by a function of the features (covariates), and where that function has a specific known form (e.g., linear separator, decision tree, union of intervals, etc.); the set of classifiers having this known form is referred to as the concept space. The natural formalization of the realizable case is very much analogous to the well-known PAC model for passive learning (Valiant, 1984). In the realizable case, there are obvious examples of learning problems where active learning can provide a significant advantage compared to passive learning; for instance, in the problem of learning threshold classifiers on the real line (Example 1 below), a kind of binary search strategy for selecting which examples to request labels for naturally leads to exponential improvements in label complexity compared to learning from random labeled examples (passive learning). As such, there is a natural attraction to determine how general this phenomenon is. This leads us to think about general-purpose learning strategies (i.e., which can be instantiated for more than merely threshold classifiers on the real line), which exhibit this binary search behavior in various special cases. The first such general-purpose strategy to emerge in the literature was a particularly elegant strategy proposed by Cohn, Atlas, and Ladner (1994), typically referred to as CAL after its discoverers (Meta-Algorithm 2 below). The strategy behind CAL is the following. The algorithm examines each example in the unlabeled pool in sequence, and if there are two classifiers in the concept space consistent with all previously-observed labels, but which disagree on the label of this next example, then the algorithm requests that label, and otherwise it does not. For this reason, below we refer to the general family of algorithms inspired by CAL as disagreement-based methods. Disagreement-based methods are sometimes referred to as “mellow” active learning, since in some sense this is the least we can expect from a reasonable active learning algorithm; it never requests the label of an example whose label it can infer from information already available, but otherwise makes no attempt to seek out particularly informative examples to request the labels of. That is, the notion of informativeness implicit in disagreement-based methods is a binary one, so that an example is either informative or not informative, but there is no further ranking of the informativeness of examples. The disagreement-based strategy is quite general, and obviously leads to algorithms that are at least reasonable, but Cohn, Atlas, and Ladner (1994) did not study the label complexity achieved by their strategy in any generality. 1472 ACTIVIZED L EARNING In a Bayesian variant of the realizable setting, Freund, Seung, Shamir, and Tishby (1997) studied an algorithm known as query by committee (QBC), which in some sense represents a Bayesian variant of CAL. However, QBC does distinguish between different levels of informativeness beyond simple disagreement, based on the amount of disagreement on a random unlabeled example. They were able to analyze the label complexity achieved by QBC in terms of a type of information gain, and found that when the information gain is lower bounded by a positive constant, the algorithm achieves a label complexity exponentially smaller than the known results for passive learning. In particular, this is the case for the threshold learning problem, and also for the problem of learning higher-dimensional (nearly balanced) linear separators when the data satisfy a certain (uniform) distribution. Below, we will not discuss this analysis further, since it is for a slightly different (Bayesian) setting. However, the results below in our present setting do have interesting implications for the Bayesian setting as well, as discussed in the recent work of Yang, Hanneke, and Carbonell (2011). The first general analysis of the label complexity of active learning in the (non-Bayesian) realizable case came in the breakthrough work of Dasgupta (2005). In that work, Dasgupta proposed a quantity, called the splitting index, to characterize the label complexities achievable by active learning. The splitting index analysis is noteworthy for several reasons. First, one can show it provides nearly tight bounds on the minimax label complexity for a given concept space and data distribution. In particular, the analysis matches the exponential improvements known to be possible for threshold classifiers, as well as generalizations to higher-dimensional homogeneous linear separators under near-uniform distributions (as first established by Dasgupta, Kalai, and Monteleoni, 2005, 2009). Second, it provides a novel notion of informativeness of an example, beyond the simple binary notion of informativeness employed in disagreement-based methods. Specifically, it describes the informativeness of an example in terms of the number of pairs of well-separated classifiers for which at least one out of each pair will be contradicted, supposing the least-favorable label. Finally, unlike any other existing work on active learning (present work included), it provides an elegant description of the trade-off between the number of label requests and the number of unlabeled examples needed by the learning algorithm. Another interesting byproduct of Dasgupta’s work is a better understanding of the nature of the improvements achievable by active learning in the general case. In particular, his work clearly illustrates the need to study the label complexity as a quantity that varies depending on the particular target concept and data distribution. We will see this issue arise in many of the examples below. Coming from a slightly different perspective, Hanneke (2007a) later analyzed the label complexity of active learning in terms of an extension of the teaching dimension (Goldman and Kearns, 1995). Related quantities were previously used by Heged¨ s (1995) and Hellerstein, Pillaipakkamu natt, Raghavan, and Wilkins (1996) to tightly characterize the number of membership queries sufficient for Exact learning; Hanneke (2007a) provided a natural generalization to the PAC learning setting. At this time, it is not clear how this quantity relates to the splitting index. From a practical perspective, in some instances it may be easier to calculate (see the work of Nowak, 2008 for a discussion related to this), though in other cases the opposite seems true. The next progress toward understanding the label complexity of active learning came in the work of Hanneke (2007b), who introduced a quantity called the disagreement coefficient (Definition 9 below), accompanied by a technique for analyzing disagreement-based active learning algorithms. In particular, implicit in that work, and made explicit in the later work of Hanneke (2011), was the first general characterization of the label complexities achieved by the original CAL strategy for 1473 H ANNEKE active learning in the realizable case, stated in terms of the disagreement coefficient. The results of the present work are direct descendants of that 2007 paper, and we will discuss the disagreement coefficient, and results based on it, in substantial detail below. Disagreement-based active learners such as CAL are known to be sometimes suboptimal relative to the splitting index analysis, and therefore the disagreement coefficient analysis sometimes results in larger label complexity bounds than the splitting index analysis. However, in many cases the label complexity bounds based on the disagreement coefficient are surprisingly good considering the simplicity of the methods. Furthermore, as we will see below, the disagreement coefficient has the practical benefit of often being fairly straightforward to calculate for a variety of learning problems, particularly when there is a natural geometric interpretation of the classifiers and the data distribution is relatively smooth. As we discuss below, it can also be used to bound the label complexity of active learning in noisy settings. For these reasons (simplicity of algorithms, ease of calculation, and applicability beyond the realizable case), subsequent work on the label complexity of active learning has tended to favor the disagreement-based approach, making use of the disagreement coefficient to bound the label complexity (Dasgupta, Hsu, and Monteleoni, 2007; Friedman, 2009; Beygelzimer, Dasgupta, and Langford, 2009; Wang, 2009; Balcan, Hanneke, and Vaughan, 2010; Hanneke, 2011; Koltchinskii, 2010; Beygelzimer, Hsu, Langford, and Zhang, 2010; Mahalanabis, 2011; Wang, 2011). A significant part of the present paper focuses on extending and generalizing the disagreement coefficient analysis, while still maintaining the relative ease of calculation that makes the disagreement coefficient so useful. In addition to many positive results, Dasgupta (2005) also pointed out several negative results, even for very simple and natural learning problems. In particular, for many problems, the minimax label complexity of active learning will be no better than that of passive learning. In fact, Balcan, Hanneke, and Vaughan (2010) later showed that, for a certain type of active learning algorithm— namely, self-verifying algorithms, which themselves adaptively determine how many label requests they need to achieve a given accuracy—there are even particular target concepts and data distributions for which no active learning algorithm of that type can outperform passive learning. Since all of the above label complexity analyses (splitting index, teaching dimension, disagreement coefficient) apply to certain respective self-verifying learning algorithms, these negative results are also reflected in all of the existing general label complexity analyses. While at first these negative results may seem discouraging, Balcan, Hanneke, and Vaughan (2010) noted that if we do not require the algorithm to be self-verifying, instead simply measuring the number of label requests the algorithm needs to find a good classifier, rather than the number needed to both find a good classifier and verify that it is indeed good, then these negative results vanish. In fact, (shockingly) they were able to show that for any concept space with finite VC dimension, and any fixed data distribution, for any given passive learning algorithm there is an active learning algorithm with asymptotically superior label complexity for every nontrivial target concept! A positive result of this generality and strength is certainly an exciting advance in our understanding of the advantages of active learning. But perhaps equally exciting are the unresolved questions raised by that work, as there are potential opportunities to strengthen, generalize, simplify, and elaborate on this result. First, note that the above statement allows the active learning algorithm to be specialized to the particular distribution according to which the (unlabeled) data are sampled, and indeed the active learning method used by Balcan, Hanneke, and Vaughan (2010) in their proof has a rather strong direct dependence on the data distribution (which cannot be removed by simply replacing some calculations with data-dependent estimators). One interesting question is whether 1474 ACTIVIZED L EARNING an alternative approach might avoid this direct distribution-dependence in the algorithm, so that the claim can be strengthened to say that the active algorithm is superior to the passive algorithm for all nontrivial target concepts and data distributions. This question is interesting both theoretically, in order to obtain the strongest possible theorem on the advantages of active learning, as well as practically, since direct access to the distribution from which the data are sampled is typically not available in practical learning scenarios. A second question left open by Balcan, Hanneke, and Vaughan (2010) regards the magnitude of the gap between the active and passive label complexities. Specifically, although they did find particularly nasty learning problems where the label complexity of active learning will be close to that of passive learning (though always better), they hypothesized that for most natural learning problems, the improvements over passive learning should typically be exponentially large (as is the case for threshold classifiers); they gave many examples to illustrate this point, but left open the problem of characterizing general sufficient conditions for these exponential improvements to be achievable, even when they are not achievable by self-verifying algorithms. Another question left unresolved by Balcan, Hanneke, and Vaughan (2010) is whether this type of general improvement guarantee might be realized by a computationally efficient active learning algorithm. Finally, they left open the question of whether such general results might be further generalized to settings that involve noisy labels. The present work picks up where Balcan, Hanneke, and Vaughan (2010) left off in several respects, making progress on each of the above questions, in some cases completely resolving the question. 1.1.2 T HE AGNOSTIC C ASE In addition to the above advances in our understanding of active learning in the realizable case, there has also been wonderful progress in making these methods robust to imperfect teachers, feature space underspecification, and model misspecification. This general topic goes by the name agnostic active learning, from its roots in the agnostic PAC model (Kearns, Schapire, and Sellie, 1994). In contrast to the realizable case, in the agnostic case, there is not necessarily a perfect classifier of a known form, and indeed there may even be label noise so that there is no perfect classifier of any form. Rather, we have a given set of classifiers (e.g., linear separators, or depth-limited decision trees, etc.), and the objective is to identify a classifier whose accuracy is not much worse than the best classifier of that type. Agnostic learning is strictly more general, and often more difficult, than realizable learning; this is true for both passive learning and active learning. However, for a given agnostic learning problem, we might still hope that active learning can achieve a given accuracy using fewer labels than required for passive learning. The general topic of agnostic active learning got its first taste of real progress from Balcan, Beygelzimer, and Langford (2006a, 2009) with the publication of the A2 (agnostic active) algorithm. This method is a noise-robust disagreement-based algorithm, which can be applied with essentially arbitrary types of classifiers under arbitrary noise distributions. It is interesting both for its effectiveness and (as with CAL) its elegance. The original work of Balcan, Beygelzimer, and Langford (2006a, 2009) showed that, in some special cases (thresholds, and homogeneous linear separators under a uniform distribution), the A2 algorithm does achieve improved label complexities compared to the known results for passive learning. Using a different type of general active learning strategy, Hanneke (2007a) found that the teaching dimension analysis (discussed above for the realizable case) can be extended beyond the realizable case, arriving at general bounds on the label complexity under arbitrary noise distributions. 1475 H ANNEKE These bounds improve over the known results for passive learning in many cases. However, the algorithm requires direct access to a certain quantity that depends on the noise distribution (namely, the noise rate, defined in Section 6 below), which would not be available in many real-world learning problems. Later, Hanneke (2007b) established a general characterization of the label complexities achieved by A2 , expressed in terms of the disagreement coefficient. The result holds for arbitrary types of classifiers (of finite VC dimension) and arbitrary noise distributions, and represents the natural generalization of the aforementioned realizable-case analysis of CAL. In many cases, this result shows improvements over the known results for passive learning. Furthermore, because of the simplicity of the disagreement coefficient, the bound can be calculated for a variety of natural learning problems. Soon after this, Dasgupta, Hsu, and Monteleoni (2007) proposed a new active learning strategy, which is also effective in the agnostic setting. Like A2 , the new algorithm is a noise-robust disagreement-based method. The work of Dasgupta, Hsu, and Monteleoni (2007) is significant for at least two reasons. First, they were able to establish a general label complexity bound for this method based on the disagreement coefficient. The bound is similar in form to the previous label complexity bound for A2 by Hanneke (2007b), but improves the dependence of the bound on the disagreement coefficient. Second, the proposed method of Dasgupta, Hsu, and Monteleoni (2007) set a new standard for computational and aesthetic simplicity in agnostic active learning algorithms. This work has since been followed by related methods of Beygelzimer, Dasgupta, and Langford (2009) and Beygelzimer, Hsu, Langford, and Zhang (2010). In particular, Beygelzimer, Dasgupta, and Langford (2009) develop a method capable of learning under an essentially arbitrary loss function; they also show label complexity bounds similar to those of Dasgupta, Hsu, and Monteleoni (2007), but applicable to a larger class of loss functions, and stated in terms of a generalization of the disagreement coefficient for arbitrary loss functions. While the above results are encouraging, the guarantees reflected in these label complexity bounds essentially take the form of (at best) constant factor improvements; specifically, in some cases the bounds improve the dependence on the noise rate factor (defined in Section 6 below), compared to the known results for passive learning. In fact, K¨ ari¨ inen (2006) showed that any a¨ a label complexity bound depending on the noise distribution only via the noise rate cannot do better than this type of constant-factor improvement. This raised the question of whether, with a more detailed description of the noise distribution, one can show improvements in the asymptotic form of the label complexity compared to passive learning. Toward this end, Castro and Nowak (2008) studied a certain refined description of the noise conditions, related to the margin conditions of Mammen and Tsybakov (1999), which are well-studied in the passive learning literature. Specifically, they found that in some special cases, under certain restrictions on the noise distribution, the asymptotic form of the label complexity can be improved compared to passive learning, and in some cases the improvements can even be exponential in magnitude; to achieve this, they developed algorithms specifically tailored to the types of classifiers they studied (threshold classifiers and boundary fragment classes). Balcan, Broder, and Zhang (2007) later extended this result to general homogeneous linear separators under a uniform distribution. Following this, Hanneke (2009a, 2011) generalized these results, showing that both of the published general agnostic active learning algorithms (Balcan, Beygelzimer, and Langford, 2009; Dasgupta, Hsu, and Monteleoni, 2007) can also achieve these types of improvements in the asymptotic form of the label complexity; he further proved general bounds on the label complexities of these methods, again based on the disagreement coefficient, which apply to arbitrary types of classifiers, and which reflect these types of improvements 1476 ACTIVIZED L EARNING (under conditions on the disagreement coefficient). Wang (2009) later bounded the label complexity of A2 under somewhat different noise conditions, in particular identifying weaker noise conditions sufficient for these improvements to be exponential in magnitude (again, under conditions on the disagreement coefficient). Koltchinskii (2010) has recently improved on some of Hanneke’s results, refining certain logarithmic factors and simplifying the proofs, using a slightly different algorithm based on similar principles. Though the present work discusses only classes of finite VC dimension, most of the above references also contain results for various types of nonparametric classes with infinite VC dimension. At present, all of the published bounds on the label complexity of agnostic active learning also apply to self-verifying algorithms. As mentioned, in the realizable case, it is typically possible to achieve significantly better label complexities if we do not require the active learning algorithm to be self-verifying, since the verification of learning may be more difficult than the learning itself (Balcan, Hanneke, and Vaughan, 2010). We might wonder whether this is also true in the agnostic case, and whether agnostic active learning algorithms that are not self-verifying might possibly achieve significantly better label complexities than the existing label complexity bounds described above. We investigate this in depth below. 1.2 Summary of Contributions In the present work, we build on and extend the above results in a variety of ways, resolving a number of open problems. The main contributions of this work can be summarized as follows. • We formally define a notion of a universal activizer, a meta-algorithm that transforms any passive learning algorithm into an active learning algorithm with asymptotically strictly superior label complexities for all nontrivial distributions and target concepts in the concept space. • We analyze the existing strategy of disagreement-based active learning from this perspective, precisely characterizing the conditions under which this strategy can lead to a universal activizer for VC classes in the realizable case. • We propose a new type of active learning algorithm, based on shatterable sets, and construct universal activizers for all VC classes in the realizable case based on this idea; in particular, this overcomes the issue of distribution-dependence in the existing results mentioned above. • We present a novel generalization of the disagreement coefficient, along with a new asymptotic bound on the label complexities achievable by active learning in the realizable case; this new bound is often significantly smaller than the existing results in the published literature. • We state new concise sufficient conditions for exponential improvements over passive learning to be achievable in the realizable case, including a significant weakening of known conditions in the published literature. • We present a new general-purpose active learning algorithm for the agnostic case, based on the aforementioned idea involving shatterable sets. • We prove a new asymptotic bound on the label complexities achievable by active learning in the presence of label noise (the agnostic case), often significantly smaller than any previously published results. 1477 H ANNEKE • We formulate a general conjecture on the theoretical advantages of active learning over passive learning in the presence of arbitrary types of label noise. 1.3 Outline of the Paper The paper is organized as follows. In Section 2, we introduce the basic notation used throughout, formally define the learning protocol, and formally define the label complexity. We also define the notion of an activizer, which is a procedure that transforms a passive learning algorithm into an active learning algorithm with asymptotically superior label complexity. In Section 3, we review the established technique of disagreement-based active learning, and prove a new result precisely characterizing the scenarios in which disagreement-based active learning can be used to construct an activizer. In particular, we find that in many scenarios, disagreement-based active learning is not powerful enough to provide the desired improvements. In Section 4, we move beyond disagreementbased active learning, developing a new type of active learning algorithm based on shatterable sets of points. We apply this technique to construct a simple 3-stage procedure, which we then prove is a universal activizer for any concept space of finite VC dimension. In Section 5, we begin by reviewing the known results for bounding the label complexity of disagreement-based active learning in terms of the disagreement coefficient; we then develop a somewhat more involved procedure, again based on shatterable sets, which takes full advantage of the sequential nature of active learning. In addition to being an activizer, we show that this procedure often achieves dramatically superior label complexities than achievable by passive learning. In particular, we define a novel generalization of the disagreement coefficient, and use it to bound the label complexity of this procedure. This also provides us with concise sufficient conditions for obtaining exponential improvements over passive learning. Continuing in Section 6, we extend our framework to allow for label noise (the agnostic case), and discuss the possibility of extending the results from previous sections to these noisy learning problems. We first review the known results for noise-robust disagreement-based active learning, and characterizations of its label complexity in terms of the disagreement coefficient and Mammen-Tsybakov noise parameters. We then proceed to develop a new type of noise-robust active learning algorithm, again based on shatterable sets, and prove bounds on its label complexity in terms of our aforementioned generalization of the disagreement coefficient. Additionally, we present a general conjecture concerning the existence of activizers for certain passive learning algorithms in the agnostic case. We conclude in Section 7 with a host of enticing open problems for future investigation. 2. Definitions and Notation For most of the paper, we consider the following formal setting. There is a measurable space (X , FX ), where X is called the instance space; for simplicity, we suppose this is a standard Borel space (Srivastava, 1998) (e.g., Rm under the usual Borel σ -algebra), though most of the results generalize. A classifier is any measurable function h : X → {−1, +1}. There is a set C of classifiers called the concept space. In the realizable case, the learning problem is characterized as follows. There is a probability measure P on X , and a sequence ZX = {X1 , X2 , . . .} of independent X -valued random variables, each with distribution P. We refer to these random variables as the sequence of unlabeled examples; although in practice, this sequence would typically be large but finite, to simplify the discussion and focus strictly on counting labels, we will suppose this sequence is inexhaustible. There is additionally a special element f ∈ C, called the target function, and we denote by 1478 ACTIVIZED L EARNING Yi = f (Xi ); we further denote by Z = {(X1 ,Y1 ), (X2 ,Y2 ), . . .} the sequence of labeled examples, and for m ∈ N we denote by Zm = {(X1 ,Y1 ), (X2 ,Y2 ), . . . , (Xm ,Ym )} the finite subsequence consisting of the first m elements of Z. For any classifier h, we define the error rate er(h) = P(x : h(x) = f (x)). Informally, the learning objective in the realizable case is to identify some h with small er(h) using elements from Z, without direct access to f . An active learning algorithm A is permitted direct access to the ZX sequence (the unlabeled examples), but to gain access to the Yi values it must request them one at a time, in a sequential manner. Specifically, given access to the ZX values, the algorithm selects any index i ∈ N, requests to observe the Yi value, then having observed the value of Yi , selects another index i′ , observes the value of Yi′ , etc. The algorithm is given as input an integer n, called the label budget, and is permitted ˆ to observe at most n labels total before eventually halting and returning a classifier hn = A(n); that is, by definition, an active learning algorithm never attempts to access more than the given budget n ˆ number of labels. We will then study the values of n sufficient to guarantee E[er(hn )] ≤ ε , for any given value ε ∈ (0, 1). We refer to this as the label complexity. We will be particularly interested in the asymptotic dependence on ε in the label complexity, as ε → 0. Formally, we have the following definition. Definition 1 An active learning algorithm A achieves label complexity Λ(·, ·, ·) if, for every target function f , distribution P, ε ∈ (0, 1), and integer n ≥ Λ(ε , f , P), we have E [er (A(n))] ≤ ε . This definition of label complexity is similar to one originally studied by Balcan, Hanneke, and Vaughan (2010). It has a few features worth noting. First, the label complexity has an explicit dependence on the target function f and distribution P. As noted by Dasgupta (2005), we need this dependence if we are to fully understand the range of label complexities achievable by active learning; we further illustrate this issue in the examples below. The second feature to note is that the label complexity, as defined here, is simply a sufficient budget size to achieve the specified accuracy. That is, here we are asking only how many label requests are required for the algorithm to achieve a given accuracy (in expectation). However, as noted by Balcan, Hanneke, and Vaughan (2010), this number might not be sufficiently large to detect that the algorithm has indeed achieved the required accuracy based only on the observed data. That is, because the number of labeled examples used in active learning can be quite small, we come across the problem that the number of labels needed to learn a concept might be significantly smaller than the number of labels needed to verify that we have successfully learned the concept. As such, this notion of label complexity is most useful in the design of effective learning algorithms, rather than for predicting the number of labels an algorithm should request in any particular application. Specifically, to design effective active learning algorithms, we should generally desire small label complexity values, so that (in the extreme case) if some algorithm A has smaller label complexity values than some other algorithm A′ for all target functions and distributions, then (all other factors being equal) we should clearly prefer algorithm A over algorithm A′ ; this is true regardless of whether we have a means to detect (verify) how large the improvements offered by algorithm A over algorithm A′ are for any particular application. Thus, in our present context, performance guarantees in terms of this notion of label complexity play a role analogous to concepts such as universal consistency or admissibility, which are also generally useful in guiding the design of effective algorithms, but are not intended to be informative in the context of any particular application. See the work of Balcan, Hanneke, and Vaughan (2010) for a discussion of this issue, as it relates to a definition of label complexity similar 1479 H ANNEKE to that above, as well as other notions of label complexity from the active learning literature (some of which include a verification requirement). We will be interested in the performance of active learning algorithms, relative to the performance of a given passive learning algorithm. In this context, a passive learning algorithm A takes as input a finite sequence of labeled examples L ∈ n (X × {−1, +1})n , and returns a classifier ˆ h = A(L). We allow both active and passive learning algorithms to be randomized: that is, to have independent internal randomness, in addition to the given random data. We define the label complexity for a passive learning algorithm as follows. Definition 2 A passive learning algorithm A achieves label complexity Λ(·, ·, ·) if, for every target function f , distribution P, ε ∈ (0, 1), and integer n ≥ Λ(ε , f , P), we have E [er (A (Zn ))] ≤ ε . Although technically some algorithms may be able to achieve a desired accuracy without any observations, to make the general results easier to state (namely, those in Section 5), unless otherwise stated we suppose label complexities (both passive and active) take strictly positive values, among N ∪ {∞}; note that label complexities (both passive and active) can be infinite, indicating that the corresponding algorithm might not achieve expected error rate ε for any n ∈ N. Both the passive and active label complexities are defined as a number of labels sufficient to guarantee the expected error rate is at most ε . It is also common in the literature to discuss the number of label requests sufficient to guarantee the error rate is at most ε with high probability 1 − δ (e.g., Balcan, Hanneke, and Vaughan, 2010). In the present work, we formulate our results in terms of the expected error rate because it simplifies the discussion of asymptotics, in that we need only study the behavior of the label complexity as the single argument ε approaches 0, rather than the more complicated behavior of a function of ε and δ as both ε and δ approach 0 at various relative rates. However, we note that analogous results for these high-probability guarantees on the error rate can be extracted from the proofs below without much difficulty, and in several places we explicitly state results of this form. Below we employ the standard notation from asymptotic analysis, including O(·), o(·), Ω(·), ω (·), Θ(·), ≪, and ≫. In all contexts below not otherwise specified, the asymptotics are always considered as ε → 0 when considering a function of ε , and as n → ∞ when considering a function of n; also, in any expression of the form “x → 0,” we always mean the limit from above (i.e., x ↓ 0). For instance, when considering nonnegative functions of ε , λa (ε ) and λ p (ε ), the above notations λ are defined as follows. We say λa (ε ) = o(λ p (ε )) when lim λa (ε ) = 0, and this is equivalent to p (ε ) ε →0 writing λ p (ε ) = ω (λa (ε )), λa (ε ) ≪ λ p (ε ), or λ p (ε ) ≫ λa (ε ). We say λa (ε ) = O(λ p (ε )) when λ lim sup λa (ε ) < ∞, which can equivalently be expressed as λ p (ε ) = Ω(λa (ε )). Finally, we write p (ε ) ε →0 λa (ε ) = Θ(λ p (ε )) to mean that both λa (ε ) = O(λ p (ε )) and λa (ε ) = Ω(λ p (ε )) are satisfied. We also use the standard notation for the limit of a sequence of sets, such as lim Ar , defined by the r→0 property ½ lim Ar = lim ½Ar (if the latter exists), where ½A is the indicator function for the set A. r→0 r→0 Define the class of functions Polylog(1/ε ) as those g : (0, 1) → [0, ∞) such that, for some k ∈ [0, ∞), g(ε ) = O(logk (1/ε )). For a label complexity Λ, also define the set Nontrivial(Λ) as the collection of all pairs ( f , P) of a classifier and a distribution such that, ∀ε > 0, Λ(ε , f , P) < ∞, and ∀g ∈ Polylog(1/ε ), Λ(ε , f , P) = ω (g(ε )). 1480 ACTIVIZED L EARNING In this context, an active meta-algorithm is a procedure Aa taking as input a passive algorithm A p and a label budget n, such that for any passive algorithm A p , Aa (A p , ·) is an active learning algorithm. We define an activizer for a given passive algorithm as follows. Definition 3 We say an active meta-algorithm Aa activizes a passive algorithm A p for a concept space C if the following holds. For any label complexity Λ p achieved by A p , the active learning algorithm Aa (A p , ·) achieves a label complexity Λa such that, for every f ∈ C and every distribution P on X with ( f , P) ∈ Nontrivial(Λ p ), there exists a constant c ∈ [1, ∞) such that Λa (cε , f , P) = o (Λ p (ε , f , P)) . In this case, Aa is called an activizer for A p with respect to C, and the active learning algorithm Aa (A p , ·) is called the Aa -activized A p . We also refer to any active meta-algorithm Aa that activizes every passive algorithm A p for C as a universal activizer for C. One of the main contributions of this work is establishing that such universal activizers do exist for any VC class C. A bit of explanation is in order regarding Definition 3. We might interpret it as follows: an activizer for A p strongly improves (in a little-o sense) the label complexity for all nontrivial target functions and distributions. Here, we seek a meta-algorithm that, when given A p as input, results in an active learning algorithm with strictly superior label complexities. However, there is a sense in which some distributions P or target functions f are trivial relative to A p . For instance, perhaps A p has a default classifier that it is naturally biased toward (e.g., with minimal P(x : h(x) = +1), as in the Closure algorithm of Helmbold, Sloan, and Warmuth, 1990), so that when this default classifier is the target function, A p achieves a constant label complexity. In these trivial scenarios, we cannot hope to improve over the behavior of the passive algorithm, but instead can only hope to compete with it. The sense in which we wish to compete may be a subject of some controversy, but the implication of Definition 3 is that the label complexity of the activized algorithm should be strictly better than every nontrivial upper bound on the label complexity of the passive algorithm. For instance, if Λ p (ε , f , P) ∈ Polylog(1/ε ), then we are guaranteed Λa (ε , f , P) ∈ Polylog(1/ε ) as well, but if Λ p (ε , f , P) = O(1), we are still only guaranteed Λa (ε , f , P) ∈ Polylog(1/ε ). This serves the purpose of defining a framework that can be studied without requiring too much obsession over small additive terms in trivial scenarios, thus focusing the analyst’s efforts toward nontrivial scenarios where A p has relatively large label complexity, which are precisely the scenarios for which active learning is truly needed. In our proofs, we find that in fact Polylog(1/ε ) can be replaced with O(log(1/ε )), giving a slightly broader definition of “nontrivial,” for which all of the results below still hold. Section 7 discusses open problems regarding this issue of trivial problems. The definition of Nontrivial(·) also only requires the activized algorithm to be effective in scenarios where the passive learning algorithm has reasonable behavior (i.e., finite label complexities); this is only intended to keep with the reduction-based style of the framework, and in fact this restriction can easily be lifted using a trick from Balcan, Hanneke, and Vaughan (2010) (aggregating the activized algorithm with another algorithm that is always reasonable). Finally, we also allow a constant factor c loss in the ε argument to Λa . We allow this to be an arbitrary constant, again in the interest of allowing the analyst to focus only on the most significant aspects of the problem; for most reasonable passive learning algorithms, we typically expect Λ p (ε , f , P) = Poly(1/ε ), in which case c can be set to 1 by adjusting the leading constant factors of 1481 H ANNEKE Λa . A careful inspection of our proofs reveals that c can always be set arbitrarily close to 1 without affecting the theorems below (and in fact, we can even get c = (1 + o(1)), a function of ε ). ˆ Throughout this work, we will adopt the usual notation for probabilities, such as P(er(h) > ε ), and as usual we interpret this as measuring the corresponding event in the (implicit) underlying probability space. In particular, we make the usual implicit assumption that all sets involved in the analysis are measurable; where this assumption does not hold, we may turn to outer probabilities, though we will not make further mention of these technical details. We will also use the notation P k (·) to represent k-dimensional product measures; for instance, for a measurable set ′ ′ ′ ′ A ⊆ X k , P k (A) = P((X1 , . . . , Xk ) ∈ A), for independent P-distributed random variables X1 , . . . , Xk . 0 = {∅} and P 0 (X 0 ) = 1. Additionally, to simplify notation, we will adopt the convention that X Throughout, we will denote by ½A (z) the indicator function for a set A, which has the value 1 when z ∈ A and 0 otherwise; additionally, at times it will be more convenient to use the bipolar indicator function, defined as ½± (z) = 2½A (z) − 1. A We will require a few additional definitions for the discussion below. For any classifier h : X → {−1, +1} and finite sequence of labeled examples L ∈ m (X × {−1, +1})m , define the empirical error rate erL (h) = |L|−1 (x,y)∈L ½{−y} (h(x)); for completeness, define er∅ (h) = 0. Also, for L = Zm , the first m labeled examples in the data sequence, abbreviate this as erm (h) = erZm (h). For any probability measure P on X , set of classifiers H, classifier h, and r > 0, define BH,P (h, r) = {g ∈ H : P (x : h(x) = g(x)) ≤ r}; when P = P, the distribution of the unlabeled examples, and P is clear from the context, we abbreviate this as BH (h, r) = BH,P (h, r); furthermore, when P = P and H = C, the concept space, and both P and C are clear from the context, we abbreviate this as B(h, r) = BC,P (h, r). Also, for any set of classifiers H, and any sequence of labeled examples L ∈ m (X × {−1, +1})m , define H[L] = {h ∈ H : erL (h) = 0}; for any (x, y) ∈ X × {−1, +1}, abbreviate H[(x, y)] = H[{(x, y)}] = {h ∈ H : h(x) = y}. We also adopt the usual definition of “shattering” used in learning theory (e.g., Vapnik, 1998). Specifically, for any set of classifiers H, k ∈ N, and S = (x1 , . . . , xk ) ∈ X k , we say H shatters S if, ∀(y1 , . . . , yk ) ∈ {−1, +1}k , ∃h ∈ H such that ∀i ∈ {1, . . . , k}, h(xi ) = yi ; equivalently, H shatters S if ∃{h1 , . . . , h2k } ⊆ H such that for each i, j ∈ {1, . . . , 2k } with i = j, ∃ℓ ∈ {1, . . . , k} with hi (xℓ ) = h j (xℓ ). To simplify notation, we will also say that H shatters ∅ if and only if H = {}. As usual, we define the VC dimension of C, denoted d, as the largest integer k such that ∃S ∈ X k shattered by C (Vapnik and Chervonenkis, 1971; Vapnik, 1998). To focus on nontrivial problems, we will only consider concept spaces C with d > 0 in the results below. Generally, any such concept space C with d < ∞ is called a VC class. 2.1 Motivating Examples Throughout this paper, we will repeatedly refer to a few canonical examples. Although themselves quite toy-like, they represent the boiled-down essence of some important distinctions between various types of learning problems. In some sense, the process of grappling with the fundamental distinctions raised by these types of examples has been a driving force behind much of the recent progress in understanding the label complexity of active learning. The first example is perhaps the most classic, and is clearly the first that comes to mind when considering the potential for active learning to provide strong improvements over passive learning. Example 1 In the problem of learning threshold classifiers, we consider X = [0, 1] and C = {hz (x) = ½± (x) : z ∈ (0, 1)}. [z,1] 1482 ACTIVIZED L EARNING There is a simple universal activizer for threshold classifiers, based on a kind of binary search. Specifically, suppose n ∈ N and that A p is any given passive learning algorithm. Consider the points in {X1 , X2 , . . . , Xm }, for m = 2n−1 , and sort them in increasing order: X(1) , X(2) , . . . , X(m) . Also initialize ℓ = 0 and u = m + 1, and define X(0) = 0 and X(m+1) = 1. Now request the label of X(i) for i = ⌊(ℓ + u)/2⌋ (i.e., the median point between ℓ and u); if the label is −1, let ℓ = i, and otherwise let u = i; repeat this (requesting this median point, then updating ℓ or u accordingly) until we have u = ℓ+1. Finally, let z = X(u) , construct the labeled sequence L = {(X1 , hz (X1 )) , . . . , (Xm , hz (Xm ))}, ˆ ˆ ˆ ˆ and return the classifier h = A p (L). Since each label request at least halves the set of integers between ℓ and u, the total number of label requests is at most log2 (m) + 1 = n. Supposing f ∈ C is the target function, this procedure maintains the invariant that f (X(ℓ) ) = −1 and f (X(u) ) = +1. Thus, once we reach u = ℓ + 1, since f is a threshold, it must be some hz with z ∈ (ℓ, u]; therefore every X( j) with j ≤ ℓ has f (X( j) ) = −1, and likewise every X( j) with j ≥ u has f (X( j) ) = +1; in particular, this means L ˆ equals Zm , the true labeled sequence. But this means h = A p (Zm ). Since n = log2 (m) + 1, this active learning algorithm will achieve an equivalent error rate to what A p achieves with m labeled examples, but using only log2 (m) + 1 label requests. In particular, this implies that if A p achieves label complexity Λ p , then this active learning algorithm achieves label complexity Λa such that Λa (ε , f , P) ≤ log2 Λ p (ε , f , P) + 2; as long as 1 ≪ Λ p (ε , f , P) < ∞, this is o(Λ p (ε , f , P)), so that this procedure activizes A p for C. The second example we consider is almost equally simple (only increasing the VC dimension from 1 to 2), but is far more subtle in terms of how we must approach its analysis in active learning. Example 2 In the problem of learning interval classifiers, we consider X = [0, 1] and C = {h[a,b] (x) = ½± (x) : 0 < a ≤ b < 1}. [a,b] For the intervals problem, we can also construct a universal activizer, though slightly more complicated. Specifically, suppose again that n ∈ N and that A p is any given passive learning algorithm. We first request the labels {Y1 ,Y2 , . . . ,Y⌈n/2⌉ } of the first ⌈n/2⌉ examples in the sequence. If every one of these labels is −1, then we immediately return the all-negative constant classifier ˆ h(x) = −1. Otherwise, consider the points {X1 , X2 , . . . , Xm }, for m = max 2⌊n/4⌋−1 , n , and sort them in increasing order X(1) , X(2) , . . . , X(m) . For some value i ∈ {1, . . . , ⌈n/2⌉} with Yi = +1, let j+ denote the corresponding index j such that X( j) = Xi . Also initialize ℓ1 = 0, u1 = ℓ2 = j+ , and u2 = m + 1, and define X(0) = 0 and X(m+1) = 1. Now if ℓ1 + 1 < u1 , request the label of X(i) for i = ⌊(ℓ1 + u1 )/2⌋ (the median point between ℓ1 and u1 ); if the label is −1, let ℓ1 = i, and otherwise let u1 = i; repeat this (requesting this median point, then updating ℓ1 or u1 accordingly) until we have u1 = ℓ1 + 1. Now if ℓ2 + 1 < u2 , request the label of X(i) for i = ⌊(ℓ2 + u2 )/2⌋ (the median point between ℓ2 and u2 ); if the label is −1, let u2 = i, and otherwise let ℓ2 = i; repeat this (requesting this median point, then updating u2 or ℓ2 accordingly) until we have u2 = ℓ2 + 1. Finally, let a = u1 and ˆ ˆ = ℓ2 , construct the labeled sequence L = X1 , h ˆ (X1 ) , . . . , Xm , h ˆ (Xm ) , and return the b [a,b] ˆ [a,b] ˆ ˆ classifier h = A p (L). Since each label request in the second phase halves the set of values between either ℓ1 and u1 or ℓ2 and u2 , the total number of label requests is at most min {m, ⌈n/2⌉ + 2 log2 (m) + 2} ≤ n. Suppose f ∈ C is the target function, and let w( f ) = P(x : f (x) = +1). If w( f ) = 0, then with ˆ ˆ probability 1 the algorithm will return the constant classifier h(x) = −1, which has er(h) = 0 in this 2 1 case. Otherwise, if w( f ) > 0, then for any n ≥ w( f ) ln ε , with probability at least 1 − ε , there exists 1483 H ANNEKE i ∈ {1, . . . , ⌈n/2⌉} with Yi = +1. Let H+ denote the event that such an i exists. Supposing this is the case, the algorithm will make it into the second phase. In this case, the procedure maintains the invariant that f (X(ℓ1 ) ) = −1, f (X(u1 ) ) = f (X(ℓ2 ) ) = +1, and f (X(u2 ) ) = −1, where ℓ1 < u1 ≤ ℓ2 < u2 . Thus, once we have u1 = ℓ1 + 1 and u2 = ℓ2 + 1, since f is an interval, it must be some h[a,b] with a ∈ (ℓ1 , u1 ] and b ∈ [ℓ2 , u1 ); therefore, every X( j) with j ≤ ℓ1 or j ≥ u2 has f (X( j) ) = −1, and likewise every X( j) with u1 ≤ j ≤ ℓ2 has f (X( j) ) = +1; in particular, this means L equals Zm , the true ˆ labeled sequence. But this means h = A p (Zm ). Supposing A p achieves label complexity Λ p , and ˆ that n ≥ max 8 + 4 log2 Λ p (ε , f , P), 2 ln 1 , then m ≥ 2⌊n/4⌋−1 ≥ Λ p (ε , f , P) and E er(h) ≤ w( f ) ε ˆ E er(h)½H+ + (1 − P(H+ )) ≤ E [er(A p (Zm ))] + ε ≤ 2ε . In particular, this means this active learning algorithm achieves label complexity Λa such that, for any f ∈ C with w( f ) = 0, Λa (2ε , f , P) = 0, 2 1 and for any f ∈ C with w( f ) > 0, Λa (2ε , f , P) ≤ max 8 + 4 log2 Λ p (ε , f , P), w( f ) ln ε . If ( f , P) ∈ 1 2 Nontrivial(Λ p ), then w( f ) ln ε = o(Λ p (ε , f , P)) and 8 + 4 log2 Λ p (ε , f , P) = o(Λ p (ε , f , P)), so that Λa (2ε , f , P) = o(Λ p (ε , f , P)). Therefore, this procedure activizes A p for C. This example also brings to light some interesting phenomena in the analysis of the label complexity of active learning. Note that unlike the thresholds example, we have a much stronger dependence on the target function in these label complexity bounds, via the w( f ) quantity. This issue is fundamental to the problem, and cannot be avoided. In particular, when P([0, x]) is continuous, this is the very issue that makes the minimax label complexity for this problem (i.e., minΛa max f ∈C Λa (ε , f , P)) no better than passive learning (Dasgupta, 2005). Thus, this problem emphasizes the need for any informative label complexity analysis of active learning to explicitly describe the dependence of the label complexity on the target function, as advocated by Dasgupta (2005). This example also highlights the unverifiability phenomenon explored by Balcan, Hanneke, and Vaughan (2010), since in the case of w( f ) = 0, the error rate of the returned classifier is zero, but (for nondegenerate P) there is no way for the algorithm to verify this fact based only on the finite number of labels it observes. In fact, Balcan, Hanneke, and Vaughan (2010) have shown that under continuous P, for any f ∈ C with w( f ) = 0, the number of labels required to both find a classifier of small error rate and verify that the error rate is small based only on observable quantities is essentially no better than for passive learning. These issues are present to a small degree in the intervals example, but were easily handled in a very natural way. The target-dependence shows up only in an initial phase of waiting for a positive example, and the always-negative classifiers were handled by setting a default return value. However, we can amplify these issues so that they show up in more subtle and involved ways. Specifically, consider the following example, studied by Balcan, Hanneke, and Vaughan (2010). Example 3 In the problem of learning unions of i intervals, we consider X = [0, 1] and C = hz (x) = ½±i j=1 [z2 j−1 ,z2 j ] (x) : 0 < z1 ≤ z2 ≤ . . . ≤ z2i < 1 . The challenge of this problem is that, because sometimes z j = z j+1 for some j values, we do not know how many intervals are required to minimally represent the target function: only that it is at most i. This issue will be made clearer below. We can essentially think of any effective strategy here as having two components: one component that searches (perhaps randomly) with the purpose of identifying at least one example from each decision region, and another component that refines our estimates of the end-points of the regions the first component identifies. Later, we will go through the behavior of a universal activizer for this problem in detail. 1484 ACTIVIZED L EARNING 3. Disagreement-Based Active Learning At present, perhaps the best-understood active learning algorithms are those choosing their label requests based on disagreement among a set of remaining candidate classifiers. The canonical algorithm of this type, a version of which we discuss below in Section 5.1, was proposed by Cohn, Atlas, and Ladner (1994). Specifically, for any set H of classifiers, define the region of disagreement: DIS(H) = {x ∈ X : ∃h1 , h2 ∈ H s.t. h1 (x) = h2 (x)} . The basic idea of disagreement-based algorithms is that, at any given time in the algorithm, there is a subset V ⊆ C of remaining candidates, called the version space, which is guaranteed to contain the target f . When deciding whether to request a particular label Yi , the algorithm simply checks whether Xi ∈ DIS(V ): if so, the algorithm requests Yi , and otherwise it does not. This general strategy is reasonable, since for any Xi ∈ DIS(V ), the label agreed upon by V must be f (Xi ), / so that we would get no information by requesting Yi ; that is, for Xi ∈ DIS(V ), we can accurately / infer Yi based on information already available. This type of algorithm has recently received substantial attention, not only for its obvious elegance and simplicity, but also because (as we discuss in Section 6) there are natural ways to extend the technique to the general problem of learning with label noise and model misspecification (the agnostic setting). The details of disagreement-based algorithms can vary in how they update the set V and how frequently they do so, but it turns out almost all disagreement-based algorithms share many of the same fundamental properties, which we describe below. 3.1 A Basic Disagreement-Based Active Learning Algorithm In Section 5.1, we discuss several known results on the label complexities achievable by these types of active learning algorithms. However, for now let us examine a very basic algorithm of this type. The following is intended to be a simple representative of the family of disagreement-based active learning algorithms. It has been stripped down to the bare essentials of what makes such algorithms work. As a result, although the gap between its label complexity and that achieved by passive learning is not necessarily as large as those achieved by the more sophisticated disagreement-based active learning algorithms of Section 5.1, it has the property that whenever those more sophisticated methods have label complexities asymptotically superior to those achieved by passive learning, that guarantee will also be true for this simpler method, and vice versa. The algorithm operates in only 2 phases. In the first, it uses one batch of label requests to reduce the version space V to a subset of C; in the second, it uses another batch of label requests, this time only requesting labels for points in DIS(V ). Thus, we have isolated precisely that aspect of disagreement-based active learning that involves improvements due to only requesting the labels of examples in the region of disagreement. ˆ The procedure is formally defined as follows, in terms of an estimator Pn (DIS(V )) specified below. 1485 H ANNEKE Meta-Algorithm 0 Input: passive algorithm A p , label budget n ˆ Output: classifier h 0. 1. 2. 3. 4. 5. 6. 7. 8. Request the first ⌊n/2⌋ labels {Y1 , . . . ,Y⌊n/2⌋ }, and let t ← ⌊n/2⌋ Let V = {h ∈ C : er⌊n/2⌋ (h) = 0} ˆ ˆ Let ∆ ← Pn (DIS(V )) Let L ← {} ˆ For m = ⌊n/2⌋ + 1, . . . ⌊n/2⌋ + ⌊n/(4∆)⌋ If Xm ∈ DIS(V ) and t < n, request the label Ym of Xm , and let y ← Ym and t ← t + 1 ˆ Else let y ← h(Xm ) for an arbitrary h ∈ V ˆ Let L ← L ∪ {(Xm , y)} ˆ Return A p (L) ˆ Meta-Algorithm 0 depends on a data-dependent estimator Pn (DIS(V )) of P(DIS(V )), which we can define in a variety of ways using only unlabeled examples. In particular, for the theorems ˆ below, we will take the following definition for Pn (DIS(V )), designed to be a confidence upper bound on P(DIS(V )). Let Un = {Xn2 +1 , . . . , X2n2 }. Then define 2 4 ˆ n (DIS(V )) = max P ½DIS(V ) (x), . (1) n2 n x∈Un Meta-Algorithm 0 is divided into two stages: one stage where we focus on reducing V , and a second stage where we construct the sample L for the passive algorithm. This might intuitively seem somewhat wasteful, as one might wish to use the requested labels from the first stage to augment those in the second stage when constructing L, thus feeding all of the observed labels into the passive algorithm A p . Indeed, this can improve the label complexity in some cases (albeit only by a constant factor); however, in order to get the general property of being an activizer for all passive algorithms A p , we construct the sample L so that the conditional distribution of the X components in L given |L| is P |L| , so that it is (conditionally) an i.i.d. sample, which is essential to our analysis. The choice of the number of (unlabeled) examples to process in the second stage guarantees (by a Chernoff bound) that the “t < n” constraint in Step 5 is redundant; this is a trick we will employ in several of the methods below. As explained above, because f ∈ V , this implies that every (x, y) ∈ L has y = f (x). To give some basic intuition for how this algorithm behaves, consider the example of learning ˆ threshold classifiers (Example 1); to simplify the explanation, for now we ignore the fact that Pn is only an estimate, as well as the “t < n” constraint in Step 5 (both of which will be addressed in the general analysis below). In this case, suppose the target function is f = hz . Let a = max{Xi : Xi < z, 1 ≤ i ≤ ⌊n/2⌋} and b = min{Xi : Xi ≥ z, 1 ≤ i ≤ ⌊n/2⌋}. Then V = {hz ′ : a < z ′ ≤ b} and DIS(V ) = (a, b), so that the second phase of the algorithm only requests labels for a number of points in the region (a, b). With probability 1 − ε , the probability mass in this region is at most O(log(1/ε )/n), so that |L| ≥ ℓn,ε = Ω(n2 / log(1/ε )); also, since the labels in L are all correct, and the Xm values in L are conditionally iid (with distribution P) given |L|, we see that the conditional distribution of L given |L| = ℓ is the same as the (unconditional) distribution of Zℓ . In particular, if ˆ A p achieves label complexity Λ p , and hn is the classifier returned by Meta-Algorithm 0 applied to 1486 ACTIVIZED L EARNING A p , then for any n = Ω ˆ E er hn Λ p (ε , f , P) log(1/ε ) chosen so that ℓn,ε ≥ Λ p (ε , f , P), we have ≤ ε + sup E [er (A p (Zℓ ))] ≤ ε + ℓ≥ℓn,ε sup ℓ≥Λ p (ε , f ,P) E [er (A p (Zℓ ))] ≤ 2ε . This indicates the active learning algorithm achieves label complexity Λa with Λa (2ε , f , P) = O Λ p (ε , f , P) log(1/ε ) . In particular, if ∞ > Λ p (ε , f , P) = ω (log(1/ε )), then Λa (2ε , f , P) = o(Λ p (ε , f , P)). Therefore, Meta-Algorithm 0 is a universal activizer for the space of threshold classifiers. In contrast, consider the problem of learning interval classifiers (Example 2). In this case, suppose the target function f has P(x : f (x) = +1) = 0, and that P is uniform in [0, 1]. Since (with probability one) every Yi = −1, we have V = {h[a,b] : {X1 , . . . , X⌊n/2⌋ } ∩ [a, b] = ∅}. But this contains classifiers h[a,a] for every a ∈ (0, 1) \ {X1 , . . . , X⌊n/2⌋ }, so that DIS(V ) = (0, 1) \ {X1 , . . . , X⌊n/2⌋ }. Thus, P(DIS(V )) = 1, and |L| = O(n); that is, A p gets run with no more labeled examples than simple passive learning would use. This indicates we should not expect Meta-Algorithm 0 to be a universal activizer for interval classifiers. Below, we formalize this by constructing a passive learning algorithm A p that Meta-Algorithm 0 does not activize in scenarios of this type. 3.2 The Limiting Region of Disagreement In this subsection, we generalize the examples from the previous subsection. Specifically, we prove that the performance of Meta-Algorithm 0 is intimately tied to a particular limiting set, referred to as the disagreement core. A similar definition was given by Balcan, Hanneke, and Vaughan (2010) (there referred to as the boundary, for reasons that will become clear below); it is also related to certain quantities in the work of Hanneke (2007b, 2011) described below in Section 5.1. Definition 4 Define the disagreement core of a classifier f with respect to a set of classifiers H and probability measure P as ∂H,P f = lim DIS (BH,P ( f , r)) . r→0 When P = P, the data distribution on X , and P is clear from the context, we abbreviate this as ∂H f = ∂H,P f ; if additionally H = C, the full concept space, which is clear from the context, we further abbreviate this as ∂ f = ∂C f = ∂C,P f . As we will see, disagreement-based algorithms often tend to focus their label requests around the disagreement core of the target function. As such, the concept of the disagreement core will be essential in much of our discussion below. We therefore go through a few examples to build intuition about this concept and its properties. Perhaps the simplest example to start with is C as the class of threshold classifiers (Example 1), under P uniform on [0, 1]. For any hz ∈ C and sufficiently small r > 0, B( f , r) = {hz ′ : |z ′ − z| ≤ r}, and DIS(B( f , r)) = [z − r, z + r). Therefore, ∂hz = lim DIS(B(hz , r)) = lim [z − r, z + r) = {z}. Thus, in this case, the disagreement core r→0 r→0 of hz with respect to C and P is precisely the decision boundary of the classifier. As a slightly more involved example, consider again the example of interval classifiers (Example 2), again under P uniform on [0, 1]. Now for any h[a,b] ∈ C with b − a > 0, for any sufficiently small r > 0, B(h[a,b] , r) = {h[a′ ,b′ ] : |a − a′ | + |b − b′ | ≤ r}, and DIS(B(h[a,b] , r)) = [a − r, a + r) ∪ (b − r, b + r]. Therefore, ∂h[a,b] = lim DIS(B(h[a,b] , r)) = lim [a − r, a + r) ∪ (b − r, b + r] = {a, b}. Thus, in this r→0 r→0 case as well, the disagreement core of h[a,b] with respect to C and P is again the decision boundary of the classifier. 1487 H ANNEKE As the above two examples illustrate, ∂ f often corresponds to the decision boundary of f in some geometric interpretation of X and f . Indeed, under fairly general conditions on C and P, the disagreement core of f does correspond to (a subset of) the set of points dividing the two label regions of f ; for instance, Friedman (2009) derives sufficient conditions, under which this is the case. In these cases, the behavior of disagreement-based active learning algorithms can often be interpreted in the intuitive terms of seeking label requests near the decision boundary of the target function, to refine an estimate of that boundary. However, in some more subtle scenarios this is no longer the case, for interesting reasons. To illustrate this, let us continue the example of interval classifiers from above, but now consider h[a,a] (i.e., h[a,b] with a = b). This time, for any r ∈ (0, 1) we have B(h[a,a] , r) = {h[a′ ,b′ ] ∈ C : b′ − a′ ≤ r}, and DIS(B(h[a,a] , r)) = (0, 1). Therefore, ∂h[a,a] = lim DIS(B(h[a,a] , r)) = lim (0, 1) = (0, 1). r→0 r→0 This example shows that in some cases, the disagreement core does not correspond to the decision boundary of the classifier, and indeed has P(∂ f ) > 0. Intuitively, as in the above example, this typically happens when the decision surface of the classifier is in some sense simpler than it could be. For instance, consider the space C of unions of two intervals (Example 3 with i = 2) under uniform P. The classifiers f ∈ C with P(∂ f ) > 0 are precisely those representable (up to probability zero differences) as a single interval. The others (with 0 < z1 < z2 < z3 < z4 < 1) have ∂hz = {z1 , z2 , z3 , z4 }. In these examples, the f ∈ C with P(∂ f ) > 0 are not only simpler than other nearby classifiers in C, but they are also in some sense degenerate relative to the rest of C; however, it turns out this is not always the case, as there exist scenarios (C, P), even with d = 2, and even with countable C, for which every f ∈ C has P(∂ f ) > 0; in these cases, every classifier is in some important sense simpler than some other subset of nearby classifiers in C. In Section 3.3, we show that the label complexity of disagreement-based active learning is intimately tied to the disagreement core. In particular, scenarios where P(∂ f ) > 0, such as those mentioned above, lead to the conclusion that disagreement-based methods are sometimes insufficient for activized learning. This motivates the design of more sophisticated methods in Section 4, which overcome this deficiency, along with a corresponding refinement of the definition of “disagreement core ” in Section 5.2 that eliminates the above issue with “simple” classifiers. 3.3 Necessary and Sufficient Conditions for Disagreement-Based Activized Learning In the specific case of Meta-Algorithm 0, for large n we may intuitively expect it to focus its second batch of label requests in and around the disagreement core of the target function. Thus, whenever P(∂ f ) = 0, we should expect the label requests to be quite focused, and therefore the algorithm should achieve smaller label complexity compared to passive learning. On the other hand, if P(∂ f ) > 0, then the label requests will not become focused beyond a constant fraction of the space, so that the improvements achieved by Meta-Algorithm 0 over passive learning should be, at best, a constant factor. This intuition is formalized in the following general theorem, the proof of which is included in Appendix A. Theorem 5 For any VC class C, Meta-Algorithm 0 is a universal activizer for C if and only if every f ∈ C and distribution P has P (∂C,P f ) = 0. While the formal proof is given in Appendix A, the general idea is simple. As we always have f ∈ V , any y inferred in Step 6 must equal f (x), so that all of the labels in L are correct. Also, as n ˆ grows large, classic results on passive learning imply the diameter of the set V will become small, 1488 ACTIVIZED L EARNING shrinking to zero as n → ∞ (Vapnik and Chervonenkis, 1971; Vapnik, 1982; Blumer, Ehrenfeucht, Haussler, and Warmuth, 1989). Therefore, as n → ∞, DIS(V ) should converge to a subset of ∂ f , ˆ so that in the case P(∂ f ) = 0, we have ∆ → 0; thus |L| ≫ n, which implies an asymptotic strict improvement in label complexity over the passive algorithm A p that L is fed into in Step 8. On the other hand, since ∂ f is defined by classifiers arbitrarily close to f , it is unlikely that any finite sample of correctly labeled examples can contradict enough classifiers to make DIS(V ) significantly smaller ˆ than ∂ f , so that we always have P(DIS(V )) ≥ P(∂ f ). Therefore, if P(∂ f ) > 0, then ∆ converges to some nonzero constant, so that |L| = O(n), representing only a constant factor improvement in label complexity. In fact, as is implied from this sketch (and is proven in Appendix A), the targets f and distributions P for which Meta-Algorithm 0 achieves asymptotic strict improvements for all passive learning algorithms (for which f and P are nontrivial) are precisely those (and only those) for which P(∂C,P f ) = 0. There are some general conditions under which the zero-probability disagreement cores condition of Theorem 5 will hold. For instance, it is not difficult to show this will always hold when X is countable. Furthermore, with some effort one can show it will hold for most classes having VC dimension one (e.g., any countable C with d = 1). However, as we have seen, not all spaces C satisfy this zero-probability disagreement cores property. In particular, for the interval classifiers studied in Section 3.2, we have P(∂h[a,a] ) = P((0, 1)) = 1. Indeed, the aforementioned special cases aside, for most nontrivial spaces C, one can construct distributions P that in some sense make C mimic the intervals problem, so that we should typically expect disagreement-based methods will not be activizers. For detailed discussions of various scenarios where the P(∂C,P f ) = 0 condition is (or is not) satisfied for various C, P, and f , see the works of Hanneke (2009b), Hanneke (2007b), Hanneke (2011), Balcan, Hanneke, and Vaughan (2010), Friedman (2009), Wang (2009) and Wang (2011). 4. Beyond Disagreement: A Basic Activizer Since the zero-probability disagreement cores condition of Theorem 5 is not always satisfied, we are left with the question of whether there could be other techniques for active learning, beyond simple disagreement-based methods, which could activize every passive learning algorithm for every VC class. In this section, we present an entirely new type of active learning algorithm, unlike anything in the existing literature, and we show that indeed it is a universal activizer for any class C of finite VC dimension. 4.1 A Basic Activizer As mentioned, the case P(∂ f ) = 0 is already handled nicely by disagreement-based methods, since the label requests made in the second stage of Meta-Algorithm 0 will become focused into a small region, and L therefore grows faster than n. Thus, the primary question we are faced with is what to do when P(∂ f ) > 0. Since (loosely speaking) we have DIS(V ) → ∂ f in Meta-Algorithm 0, P(∂ f ) > 0 corresponds to scenarios where the label requests of Meta-Algorithm 0 will not become focused beyond a certain extent; specifically, as we show in Appendix B (Lemmas 35 and 36), P(DIS(V ) ⊕ ∂ f ) → 0 almost surely, where ⊕ is the symmetric difference, so that we expect MetaAlgorithm 0 will request labels for at least some constant fraction of the examples in L. On the one hand, this is definitely a major problem for disagreement-based methods, since it prevents them from improving over passive learning in those cases. On the other hand, if we do not 1489 H ANNEKE restrict ourselves to disagreement-based methods, we may actually be able to exploit properties of this scenario, so that it works to our advantage. In particular, in addition to the fact that P(DIS(V ) ⊕ ∂C f ) → 0, we show in Appendix B (Lemma 35) that P(∂V f ⊕ ∂C f ) = 0 (almost surely) in MetaAlgorithm 0; this implies that for sufficiently large n, a random point x1 in DIS(V ) is likely to be in ∂V f . We can exploit this fact by using x1 to split V into two subsets: V [(x1 , +1)] and V [(x1 , −1)]. Now, if x1 ∈ ∂V f , then (by definition of the disagreement core) inf er(h) = inf er(h) = h∈V [(x1 ,+1)] h∈V [(x1 ,−1)] 0. Therefore, for almost every point x ∈ DIS(V [(x1 , +1)]), the label agreed upon for x by classifiers / in V [(x1 , +1)] should be f (x). Likewise, for almost every point x ∈ DIS(V [(x1 , −1)]), the label / agreed upon for x by classifiers in V [(x1 , −1)] should be f (x). Thus, we can accurately infer the label of any point x ∈ DIS(V [(x1 , +1)]) ∩ DIS(V [(x1 , −1)]) (except perhaps a zero-probability subset). / With these sets V [(x1 , +1)] and V [(x1 , −1)] in hand, there is no longer a need to request the labels of points for which either of them has agreement about the label, and we can focus our label requests to the region DIS(V [(x1 , +1)]) ∩ DIS(V [(x1 , −1)]), which may be much smaller than DIS(V ). Now if P(DIS(V [(x1 , +1)]) ∩ DIS(V [(x1 , −1)])) → 0, then the label requests will become focused to a shrinking region, and by the same reasoning as for Theorem 5 we can asymptotically achieve strict improvements over passive learning by a method analogous to Meta-Algorithm 0 (with the above changes). Already this provides a significant improvement over disagreement-based methods in many cases; indeed, in some cases (such as intervals) this fully addresses the nonzero-probability disagreement core issue in Theorem 5. In other cases (such as unions of two intervals), it does not completely address the issue, since for some targets we do not have P(DIS(V [(x1 , +1)]) ∩ DIS(V [(x1 , −1)])) → 0. However, by repeatedly applying this same reasoning, we can address the issue in full generality. Specifically, if P(DIS(V [(x1 , +1)]) ∩ DIS(V [(x1 , −1)])) 0, then DIS(V [(x1 , +1)]) ∩ DIS(V [(x1 , −1)]) essentially converges to a region ∂C[(x1 ,+1)] f ∩ ∂C[(x1 ,−1)] f , which has nonzero probability, and is nearly equivalent to ∂V [(x1 ,+1)] f ∩ ∂V [(x1 ,−1)] f . Thus, for sufficiently large n, a random x2 in DIS(V [(x1 , +1)]) ∩ DIS(V [(x1 , −1)]) will likely be in ∂V [(x1 ,+1)] f ∩ ∂V [(x1 ,−1)] f . In this case, we can repeat the above argument, this time splitting V into four sets (V [(x1 , +1)][(x2 , +1)], V [(x1 , +1)][(x2 , −1)], V [(x1 , −1)][(x2 , +1)], and V [(x1 , −1)][(x2 , −1)]), each with infimum error rate equal zero, so that for a point x in the region of agreement of any of these four sets, the agreed-upon label will (almost surely) be f (x), so that we can infer that label. Thus, we need only request the labels of those points in the intersection of all four regions of disagreement. We can further repeat this process as many times as needed, until we get a partition of V with shrinking probability mass in the intersection of the regions of disagreement, which (as above) can then be used to obtain asymptotic improvements over passive learning. Note that the above argument can be written more concisely in terms of shattering. That is, any x ∈ DIS(V ) is simply an x such that V shatters {x}; a point x ∈ DIS(V [(x1 , +1)]) ∩ DIS(V [(x1 , −1)]) is simply one for which V shatters {x1 , x}, and for any x ∈ DIS(V [(x1 , +1)]) ∩ DIS(V [(x1 , −1)]), the / label y we infer about x has the property that the set V [(x, −y)] does not shatter {x1 }. This continues for each repetition of the above idea, with x in the intersection of the four regions of disagreement simply being one for which V shatters {x1 , x2 , x}, and so on. In particular, this perspective makes it clear that we need only repeat this idea at most d times to get a shrinking intersection region, since no set of d + 1 points is shatterable. Note that there may be unobservable factors (e.g., the target function) determining the appropriate number of iterations of this idea sufficient to have a shrinking probability of requesting a label, while maintaining the accuracy of inferred labels. To address this, 1490 ACTIVIZED L EARNING we can simply try all d + 1 possibilities, and then select one of the resulting d + 1 classifiers via a kind of tournament of pairwise comparisons. Also, in order to reduce the probability of a mistaken inference due to x1 ∈ ∂V f (or similarly for later xi ), we can replace each single xi with multiple / samples, and then take a majority vote over whether to infer the label, and which label to infer if we do so; generally, we can think of this as estimating certain probabilities, and below we write ˆ these estimators as Pm , and discuss the details of their implementation later. Combining MetaAlgorithm 0 with the above reasoning motivates a new type of active learning algorithm, referred to as Meta-Algorithm 1 below, and stated as follows. Meta-Algorithm 1 Input: passive algorithm A p , label budget n ˆ Output: classifier h 0. Request the first mn = ⌊n/3⌋ labels, {Y1 , . . . ,Ymn }, and let t ← mn 1. Let V = {h ∈ C : ermn (h) = 0} 2. For k = 1, 2, . . . , d + 1 ˆ ˆ ˆ 3. ∆(k) ← Pmn x : P S ∈ X k−1 : V shatters S ∪ {x}|V shatters S ≥ 1/2 4. Let Lk ← {} ˆ 5. For m = mn + 1, . . . , mn + ⌊n/(6 · 2k ∆(k) )⌋ ˆ m S ∈ X k−1 : V shatters S ∪ {Xm }|V shatters S ≥ 1/2 and t < ⌊2n/3⌋ 6. If P 7. Request the label Ym of Xm , and let y ← Ym and t ← t + 1 ˆ ˆ m S ∈ X k−1 :V [(Xm , −y)] does not shatter S|V shatters S 8. Else, let y ← argmax P ˆ y∈{−1,+1} 9. Let Lk ← Lk ∪ {(Xm , y)} ˆ 10. Return ActiveSelect({A p (L1 ), A p (L2 ), . . . , A p (Ld+1 )}, ⌊n/3⌋, {Xmn +maxk |Lk |+1 , . . .}) Subroutine: ActiveSelect Input: set of classifiers {h1 , h2 , . . . , hN }, label budget m, sequence of unlabeled examples U ˆ Output: classifier h 0. 1. 2. 3. 4. For each j, k ∈ {1, 2, . . . , N} s.t. j < k, m Let R jk be the first j(N− j) ln(eN) points in U ∩{x : h j (x) = hk (x)} (if such values exist) Request the labels for R jk and let Q jk be the resulting set of labeled examples Let mk j = erQ jk (hk ) ˆ Return hk , where k = max k ∈ {1, . . . , N} : max j 2 er(h ∗∗ ). In particular, this implies er(h |{x : Now suppose j ∈ {k j j k hk∗∗ (x) = h j (x)}) > 2/3 and P(x : h j (x) = hk∗∗ (x)) > 0, which again means (with probability one) |{XM , XM+1 , . . .} ∩ {x : h j (x) = hk∗∗ (x)}| ≥ Mk∗∗ . By Hoeffding’s inequality, we have that P m jk∗∗ ≤ 7/12 ≤ exp {−Mk∗∗ /72} ≤ exp {1 − m/ (72k∗ N ln(eN))} . By a union bound, we have that P ∃ j > k∗∗ : er(h j ) > 2 er(hk∗∗ ) and m jk∗∗ ≤ 7/12 ≤ (N − k∗∗ ) · exp {1 − m/ (72k∗ N ln(eN))} . ˆ In particular, when k ≥ k∗∗ , and m jk∗∗ > 7/12 for all j > k∗∗ with er(h j ) > 2 er(hk∗∗ ), it must be true that er(hk ) ≤ 2 er(hk∗∗ ) ≤ 2 er(hk∗ ). ˆ ˆ So, by a union bound, with probability ≥ 1 − eN · exp {−m/ (72k∗ N ln(eN))}, the k chosen by ActiveSelect has er(hk ) ≤ 2 er(hk∗ ). ˆ ⋆ The next two lemmas describe the limiting behavior of S k (Vm ). In particular, we see that its k limiting value is precisely ∂C f (up to zero-probability differences). Lemma 35 establishes that k (V ⋆ ) does not decrease below ∂ k f (except for a zero-probability set), and Lemma 36 establishes S m C k that its limit is not larger than ∂C f (again, except for a zero-probability set). Lemma 35 There is an event H ′ with P(H ′ ) = 1 such that on H ′ , ∀m ∈ N, ∀k ∈ {0, . . . , d˜f − 1}, for ⋆ any H with Vm ⊆ H ⊆ C, k k k P k S k (H) ∂C f = P k ∂H f ∂C f = 1, and ∀i ∈ N, ½∂ k Hf (k+1) Si = ½∂ k f Si (k+1) C . (k) k k Also, on H ′ , every such H has P k ∂H f = P k ∂C f , and Mℓ (H) → ∞ as ℓ → ∞. ⋆ Proof We will show the first claim for the set Vm , and the result will then hold for H by monotonicity. In particular, we will show this for any fixed k ∈ {0, . . . , d˜f − 1} and m ∈ N, and the k ⋆ existence of H ′ then holds by a union bound. Fix any set S ∈ ∂C f . Suppose BVm ( f , r) does not (i) (i) (i) shatter S for some r > 0. There is an infinite sequence of sets {{h1 , h2 , . . . , h2k }}i with ∀ j ≤ 2k , (i) (i) (i) ⋆ P(x : h j (x) = f (x)) ↓ 0, such that each {h1 , . . . , h2k } ⊆ B( f , r) and shatters S. Since BVm ( f , r) does not shatter S, / ⋆ 1 = inf ½ ∃ j : h j ∈ BVm ( f , r) = inf ½ ∃ j ≤ 2k , ℓ ≤ m : h j (Xℓ ) = f (Xℓ ) . (i) i (i) i 1536 ACTIVIZED L EARNING But P inf ½ ∃ j ≤ 2k , ℓ ≤ m : h j (Xℓ ) = f (Xℓ ) = 1 ≤ inf P ∃ j ≤ 2k , ℓ ≤ m : h j (Xℓ ) = f (Xℓ ) (i) (i) i i ≤ lim i→∞ (i) (i) mP x : h j (x) = f (x) = j≤2k j≤2k m lim P x : h j (x) = f (x) = 0, i→∞ ⋆ where the second inequality follows by a union bound. Therefore, ∀r > 0, P S ∈ S k BVm ( f , r) = / ¯ ⋆ 0. Furthermore, since S k BVm ( f , r) is monotonic in r, the dominated convergence theorem gives us that ⋆ P S ∈ ∂Vm f = E lim ½S k (BV ⋆ ( f ,r)) (S) = lim P S ∈ S k BVm ( f , r) / k⋆ / ¯ r→0 r→0 m = 0. ⋆ This implies that (letting S ∼ P k be independent from Vm ) k k ¯k⋆ ¯k⋆ P P k ∂Vm f ∂C f > 0 = P P k ∂Vm f ∩ ∂C f > 0 k ¯k⋆ = lim P P k ∂Vm f ∩ ∂C f > ξ ξ →0 ≤ lim 1 ξ →0 ξ k ¯k⋆ E P k ∂Vm f ∩ ∂C f 1 E ξ →0 ξ = lim ½∂C f (S)P S ∈ ∂Vm f S / k⋆ k (Markov) (Fubini) = lim 0 = 0. ξ →0 ⋆ This establishes the first claim for Vm , on an event of probability 1, and monotonicity extends the ⋆ claim to any H ⊇ Vm . Also note that, on this event, k k k k k k k P k ∂H f ≥ P k ∂H f ∩ ∂C f = P k ∂H f ∂C f P k ∂C f = P k ∂C f , k k where the last equality follows from the first claim. Noting that for H ⊆ C, ∂H f ⊆ ∂C f , we must have k k P k ∂H f = P k ∂C f . This establishes the third claim. From the first claim, for any given value of i ∈ N the second claim (k+1) ⋆ holds for Si (with H = Vm ) on an additional event of probability 1; taking a union bound over (k) all i ∈ N extends this claim to every Si on an event of probability 1. Monotonicity then implies ½∂C f Si(k+1) = ½∂V ⋆ f Si(k+1) ≤ ½∂H f Si(k+1) ≤ ½∂C f Si(k+1) , k k k k m k extending the result to general H. Also, as k < d˜f , we know P k ∂C f > 0, and since we also know ⋆ is independent from W , the strong law of large numbers implies the final claim (for V ⋆ ) on an Vm 2 m ⋆ additional event of probability 1; again, monotonicity extends this claim to any H ⊇ Vm . Intersecting the above events over values m ∈ N and k < d˜f gives the event H ′ , and as each of the above events has probability 1 and there are countably many such events, a union bound implies P(H ′ ) = 1. 1537 H ANNEKE ⋆ Note that one specific implication of Lemma 35, obtained by taking k = 0, is that on H ′ , Vm = ∅ 0 f = X 0 so that P 0 ∂ 0 f = 1, (even if f ∈ cl(C) \ C). This is because, for f ∈ cl(C), we have ∂C C ⋆ 0 0 which means P 0 ∂Vm f = 1 (on H ′ ), so that we must have ∂Vm f = X 0 , which implies Vm = ∅. In ⋆ ⋆ ⋆ particular, this also means f ∈ cl (Vm ). Lemma 36 There is a monotonic function q(r) = o(1) (as r → 0) such that, on event H ′ , for any ⋆ k ∈ 0, . . . , d˜f − 1 , m ∈ N, r > 0, and set H such that Vm ⊆ H ⊆ B( f , r), ¯k P k ∂C f S k (H) ≤ q(r). In particular, for τ ∈ N and δ > 0, on Hτ (δ ) ∩ H ′ (where Hτ (δ ) is from Lemma 29), every m ≥ τ ⋆ ¯k and k ∈ 0, . . . , d˜f − 1 has P k ∂C f S k (Vm ) ≤ q(φ (τ ; δ )). Proof Fix any k ∈ 0, . . . , d˜f − 1 . By Lemma 35, we know that on event H ′ , ¯k Pk P k ∂C f ∩ S k (H) ¯k ≤ P k ∂C f S k (H) = P k (S k (H)) ¯k P k ∂C f ∩ S k (H) Pk = ≤ k P k ∂C f ¯k ∂C f ∩ S k (H) k P k ∂H f ¯ ∂ k f ∩ S k (B ( f , r)) C k P k ∂C f . ¯k Define qk (r) as this latter quantity. Since P k ∂C f ∩ S k (B( f , r)) is monotonic in r, ¯k P k ∂C f ∩ lim S k (B( f , r)) ¯k P k ∂C f ∩ S k (B( f , r)) r→0 = lim k k ∂k f r→0 P P k ∂C f C = k ¯k P k ∂C f ∩ ∂C f = 0. k P k ∂C f This proves qk (r) = o(1). Defining q(r) = max qk (r) : k ∈ 0, 1, . . . , d˜f − 1 = o(1) completes the proof of the first claim. ⋆ For the final claim, simply recall that by Lemma 29, on Hτ (δ ), every m ≥ τ has Vm ⊆ Vτ⋆ ⊆ B( f , φ (τ ; δ )). Lemma 37 For ζ ∈ (0, 1), define rζ = sup {r ∈ (0, 1) : q(r) < ζ } /2. ⋆ On H ′ , ∀k ∈ 0, . . . , d˜f − 1 , ∀ζ ∈ (0, 1), ∀m ∈ N, for any set H such that Vm ⊆ H ⊆ B( f , rζ ), ¯ P x : P k S k (H[(x, f (x))]) S k (H) > ζ k ¯ = P x : P k S k (H[(x, f (x))]) ∂H f > ζ = 0. (16) In particular, for δ ∈ (0, 1), defining τ (ζ ; δ ) = min τ ∈ N : sup φ (m; δ ) ≤ rζ , ∀τ ≥ τ (ζ ; δ ), and ⋆ ∀m ≥ τ , on Hτ (δ ) ∩ H ′ , (16) holds for H = Vm . 1538 m≥τ ACTIVIZED L EARNING ¯k Proof Fix k, m, H as described above, and suppose q = P k ∂C f |S k (H) < ζ ; by Lemma 36, this ′ . Since, ∂ k f ⊆ S k (H), we have that ∀x ∈ X , happens on H H k k ¯ ¯ P k S k (H[(x, f (x))]) S k (H) = P k S k (H[(x, f (x))]) ∂H f P k ∂H f S k (H) ¯k ¯k ¯ + P k S k (H[(x, f (x))]) S k (H) ∩ ∂H f P k ∂H f S k (H) . Since all probability values are bounded by 1, we have k ¯k ¯ ¯ P k S k (H[(x, f (x))]) S k (H) ≤ P k S k (H[(x, f (x))]) ∂H f + P k ∂H f S k (H) . (17) Isolating the right-most term in (17), by basic properties of probabilities we have ¯k P k ∂H f S k (H) k k ¯k ¯k ¯k ¯k = P k ∂H f S k (H) ∩ ∂C f P k ∂C f S k (H) + P k ∂H f S k (H) ∩ ∂C f P k ∂C f S k (H) k ¯k ¯k ≤ P k ∂C f S k (H) + P k ∂H f S k (H) ∩ ∂C f . (18) By assumption, the left term in (18) equals q. Examining the right term in (18), we see that k k k ¯k ¯k P k ∂H f S k (H) ∩ ∂C f = P k S k (H) ∩ ∂H f ∂C f /P k S k (H) ∂C f k k k ¯k ≤ P k ∂H f ∂C f /P k ∂H f ∂C f . (19) By Lemma 35, on H ′ the denominator in (19) is 1 and the numerator is 0. Thus, combining this fact with (17) and (18), we have that on H ′ , k ¯ ¯ P x : P k S k (H[(x, f (x))]) S k (H) > ζ ≤ P x : P k S k (H[(x, f (x))]) ∂H f > ζ − q . (20) Note that proving the right side of (20) equals zero will suffice to establish the result, since it upper bounds both the first expression of (16) (as just established) and the second expression of (16) (by monotonicity of measures). Letting X ∼ P be independent from the other random variables (Z,W1 ,W2 ), by Markov’s inequality, the right side of (20) is at most 1 k ¯ E P k S k (H[(X, f (X))]) ∂H f ζ −q H = k ¯ E P k S k (H[(X, f (X))]) ∩ ∂H f k (ζ − q)P k ∂H f H , and by Fubini’s theorem, this is (letting S ∼ P k be independent from the other random variables) E / ½∂H f (S)P x : S ∈ S k (H[(x, f (x))]) H k k (ζ − q)P k ∂H f . Lemma 35 implies this equals E ½∂H f (S)P x : S ∈ S k (H[(x, f (x))]) H / k k (ζ − q)P k ∂C f 1539 . (21) H ANNEKE (i) k For any fixed S ∈ ∂H f , there is an infinite sequence of sets (i) (i) (i) 2k , P x : h j (x) = f (x) ↓ 0, such that each h1 , . . . , h2k not shatter S, then (i) (i) h1 , h2 , . . . , h2k i∈N with ∀ j ≤ ⊆ H and shatters S. If H[(x, f (x))] does 1 = inf ½ ∃ j : h j ∈ H[(x, f (x))] = inf ½ ∃ j : h j (x) = f (x) . / (i) (i) i i In particular, P x : S ∈ S k (H[(x, f (x))]) ≤ P x : inf ½ ∃ j : h j (x) = f (x) = 1 / (i) i (i) (i) =P ≤ inf P x : ∃ j s.t. h j (x) = f (x) x : ∃ j : h j (x) = f (x) i i (i) (i) ≤ lim i→∞ j≤2k P x : h j (x) = f (x) = j≤2k lim P x : h j (x) = f (x) = 0. i→∞ Thus (21) is zero, which establishes the result. ⋆ The final claim is then implied by Lemma 29 and monotonicity of Vm in m: that is, on Hτ (δ ), ⋆ Vm ⊆ Vτ⋆ ⊆ B( f , φ (τ ; δ )) ⊆ B( f , rζ ). Lemma 38 For any ζ ∈ (0, 1), there are values n ∈ N and ε > 0, on event H⌊n/3⌋ ˜ (ε /2) ∩ H ′ , (ζ ) ∆n (ε ) : n ∈ N, ε ∈ (0, 1) letting V ˜ such that, for any ⋆ = V⌊n/3⌋ , (ζ ) ˜ ˜ P x : P d f −1 S ∈ X d f −1 : S ∪ {x} ∈ S d f (V ) S d f −1 (V ) ≥ ζ ≤ ∆n (ε ), (ζ ) and for any N-valued N(ε ) = ω (log(1/ε )), ∆N(ε ) (ε ) = o(1). Proof Throughout, we suppose the event H⌊n/3⌋ (ε /2) ∩ H ′ , and fix some ζ ∈ (0, 1). We have ∀x, ˜ ˜ ˜ ˜ P d f −1 S ∈ X d f −1 : S ∪ {x} ∈ S d f (V ) S d f −1 (V ) ˜ ˜ ˜ ˜ d˜ −1 f P d f −1 ∂Cf ˜ ˜ ˜ ˜ d˜ −1 ˜ ˜ ¯ d −1 f P d f −1 ∂Cf f S d f −1 (V ) ˜ ˜ ˜ ˜ d˜ −1 ˜ = P d f −1 S ∈ X d f −1 : S ∪ {x} ∈ S d f (V ) S d f −1 (V ) ∩ ∂Cf ¯ + P d f −1 S ∈ X d f −1 : S ∪ {x} ∈ S d f (V ) S d f −1 (V ) ∩ ∂Cf ≤ P d f −1 S ∈ X d f −1 : S ∪ {x} ∈ S d f (V ) S d f −1 (V ) ∩ ∂Cf d˜ −1 ˜ ˜ f S d f −1 (V ) ˜ ˜ ˜ ¯ d −1 f +P d f −1 ∂Cf f S d f −1 (V ) . (22) By Lemma 35, the left term in (22) equals ˜ ˜ d˜ −1 ˜ ˜ P d f −1 S ∈ X d f −1 : S ∪ {x} ∈ S d f (V ) S d f −1 (V ) ∩ ∂Cf ˜ ˜ ˜ d˜ −1 = P d f −1 S ∈ X d f −1 : S ∪ {x} ∈ S d f (V ) ∂Cf 1540 f , ˜ ˜ d˜ −1 f P d f −1 S d f −1 (V ) ∂Cf f ACTIVIZED L EARNING and by Lemma 36, the right term in (22) is at most q(φ (⌊n/3⌋; ε /2)). Thus, we have ˜ ˜ ˜ ˜ P x : P d f −1 S ∈ X d f −1 : S ∪ {x} ∈ S d f (V ) S d f −1 (V ) ≥ ζ ˜ ˜ ˜ d˜ −1 ≤ P x : P d f −1 S ∈ X d f −1 : S ∪ {x} ∈ S d f (V ) ∂Cf f ≥ ζ − q(φ (⌊n/3⌋; ε /2)) . (23) (ζ ) For n < 3τ (ζ /2; ε /2) (for τ (·; ·) defined in Lemma 37), we define ∆n (ε ) = 1. Otherwise, suppose n ≥ 3τ (ζ /2; ε /2), so that q(φ (⌊n/3⌋; ε /2)) < ζ /2, and thus (23) is at most ˜ ˜ ˜ d˜ −1 P x : P d f −1 S ∈ X d f −1 : S ∪ {x} ∈ S d f (V ) ∂Cf f ≥ ζ /2 . By Lemma 29, this is at most ˜ ˜ d˜ −1 ˜ P x : P d f −1 S ∈ X d f −1 : S ∪ {x} ∈ S d f (B( f , φ (⌊n/3⌋; ε /2))) ∂Cf f ≥ ζ /2 . Letting X ∼ P, by Markov’s inequality this is at most 2 ˜ ˜ ˜ d˜ −1 E P d f −1 S ∈ X d f −1 : S ∪ {X} ∈ S d f (B( f , φ (⌊n/3⌋; ε /2))) ∂Cf f ζ 2 d˜f ˜ ˜ d˜ −1 = P S ∪ {x} ∈ X d f : S ∪ {x} ∈ S d f (B( f , φ (⌊n/3⌋; ε /2))) and S ∈ ∂Cf f ˜f ζδ 2 d˜f ˜ ≤ P S d f (B( f , φ (⌊n/3⌋; ε /2))) . ˜f ζδ (24) (ζ ) Thus, defining ∆n (ε ) as (24) for n ≥ 3τ (ζ /2; ε /2) establishes the first claim. It remains only to prove the second claim. Let N(ε ) = ω (log(1/ε )). Since τ (ζ /2; ε /2) ≤ 4 4 d ln r4e + ln ε = O(log(1/ε )), we have that for all sufficiently small ε > 0, N(ε ) ≥ r ζ /2 ζ /2 (ζ ) ˜ 3τ (ζ /2; ε /2), so that ∆N(ε ) (ε ) equals (24) (with n = N(ε )). Furthermore, since δ f > 0, while ˜ d˜ P d f ∂Cf f = 0, and φ (⌊N(ε )/3⌋; ε /2) = o(1), by continuity of probability measures we know (ζ ) (24) is o(1) when n = N(ε ), so that we generally have ∆N(ε ) (ε ) = o(1). For any m ∈ N, define ˜ ˜ M(m) = m3 δ f /2. Lemma 39 There is a (C, P, f )-dependent constant c(i) ∈ (0, ∞) such that, for any τ ∈ N there is (i) an event Hτ ⊆ H ′ with (i) ˜ P Hτ ≥ 1 − c(i) · exp −M(τ )/4 (i) such that on Hτ , if d˜f ≥ 2, then ∀k ∈ 2, . . . , d˜f , ∀m ≥ τ , ∀ℓ ∈ N, for any set H such that Vℓ⋆ ⊆ H ⊆ C, (k) ˜ Mm (H) ≥ M(m). 1541 H ANNEKE Proof On H ′ , Lemma 35 implies every ½S k−1 (H) Si(k) ≥ ½∂H f Si(k) = ½∂C f Si(k) , so we k−1 k−1 (k) k−1 ˜ Si : i ≤ m3 ∩ ∂C f ≥ M(m) on an appropriate event. We know focus on showing P ∀k ∈ 2, . . . , d˜f , ∀m ≥ τ , (k) k−1 ˜ Si : i ≤ m3 ∩ ∂C f ≥ M(m) (k) k−1 ˜ Si : i ≤ m3 ∩ ∂C f < M(m) = 1 − P ∃k ∈ 2, . . . , d˜f , m ≥ τ : d˜f ≥ 1− (k) k−1 ˜ Si : i ≤ m3 ∩ ∂C f < M(m) , P m≥τ k=2 where the last line follows by a union bound. Thus, we will focus on bounding d˜f P m≥τ k=2 (k) k−1 ˜ Si : i ≤ m3 ∩ ∂C f < M(m) . Fix any k ∈ 2, . . . , d˜f , and integer m ≥ τ . Since E (k) k−1 ˜ = P k−1 ∂C f m3 ≥ δ f m3 , k−1 Si : i ≤ m3 ∩ ∂C f a Chernoff bound implies that P (k) k−1 k−1 ˜ Si : i ≤ m3 ∩ ∂C f < M(m) ≤ exp −m3 P k−1 ∂C f /8 ˜ ≤ exp −m3 δ f /8 . Thus, we have that (25) is at most d˜f m≥τ k=2 ˜ exp −m3 δ f /8 ≤ m≥τ ˜ d˜f · exp −m3 δ f /8 ≤ ˜ ≤ d˜f · exp −M(τ )/4 + d˜f · ∞ τ3 m≥τ 3 ˜ d˜f · exp −mδ f /8 ˜ exp −xδ f /8 dx ˜ ˜ = d˜f · 1 + 8/δ f · exp −M(τ )/4 ˜ ˜ ≤ 9d˜f /δ f · exp −M(τ )/4 . Note that since P(H ′ ) = 1, defining (i) Hτ = ∀k ∈ 2, . . . , d˜f , ∀m ≥ τ , (k) k−1 ˜ Si : i ≤ m3 ∩ ∂C f ≥ M(m) ∩ H ′ has the required properties. 1542 (25) ACTIVIZED L EARNING (i) Lemma 40 For any τ ∈ N, there is an event Gτ with (i) (i) P Hτ \ Gτ ˜ ˜ ≤ 121d˜f /δ f · exp −M(τ )/60 (i) such that, on Gτ , if d˜f ≥ 2, then for every integer s ≥ τ and k ∈ 2, . . . , d˜f , ∀r ∈ 0, r1/6 , (k) Ms (B ( f , r)) ≤ (3/2) (k) k−1 Si : i ≤ s3 ∩ ∂C f . (k) ˆ Proof Fix integers s ≥ τ and k ∈ 2, . . . , d˜f , and let r = r1/6 . Define the set S k−1 = Si : i ≤ s3 ∩ (k) ˆ ˆ S k−1 (B ( f , r)). Note S k−1 = Ms (B ( f , r)) and the elements of S k−1 are conditionally i.i.d. given (k) (k) Ms (B ( f , r)), each with conditional distribution equivalent to the conditional distribution of S1 (k) given S1 ∈ S k−1 (B ( f , r)) . In particular, (k) (k) k−1 k−1 ˆ E S k−1 ∩ ∂C f Ms (B ( f , r)) = P k−1 ∂C f S k−1 (B ( f , r)) Ms (B ( f , r)) . Define the event (i) Gτ (k, s) = k−1 ˆ ˆ S k−1 ≤ (3/2) S k−1 ∩ ∂C f . By Lemma 36 (indeed by definition of q(r) and r1/6 ) we have (k) (i) 1 − P Gτ (k, s) Ms (B ( f , r)) =P (k) (k) k−1 ˆ S k−1 ∩ ∂C f < (2/3)Ms (B ( f , r)) Ms (B ( f , r)) ≤P (k) (k) k−1 ˆ S k−1 ∩ ∂C f < (4/5) (1 − q (r)) Ms (B ( f , r)) Ms (B ( f , r)) ≤P (k) (k) k−1 k−1 ˆ S k−1 ∩ ∂C f < (4/5)P k−1 ∂C f S k−1 (B ( f , r)) Ms (B ( f , r)) Ms (B ( f , r)) . (26) By a Chernoff bound, (26) is at most (k) k−1 exp −Ms (B ( f , r)) P k−1 ∂C f S k−1 (B ( f , r)) /50 (k) (k) ≤ exp −Ms (B ( f , r)) (1 − q (r)) /50 ≤ exp −Ms (B ( f , r)) /60 . Thus, by Lemma 39, (i) (i) P Hτ \ Gτ (k, s) ≤ P =E (i) (k) (i) ˜ Ms (B ( f , r)) ≥ M(s) \ Gτ (k, s) (k) 1 − P Gτ (k, s) Ms (B ( f , r)) (k) ≤ E exp −Ms (B ( f , r)) /60 ½[M(s),∞) Ms(k) (B ( f , r)) ˜ ½[M(s),∞) Ms(k) (B ( f , r)) ˜ 1543 ˜ ≤ exp −M(s)/60 . H ANNEKE (i) Now defining Gτ = (i) d˜f (i) k=2 Gτ (k, s), s≥τ (i) P Hτ \ Gτ ≤ s≥τ a union bound implies ˜ d˜f · exp −M(s)/60 ˜ ≤ d˜f exp −M(τ )/60 + ∞ τ3 ˜ exp −xδ f /120 dx ˜ ˜ = d˜f 1 + 120/δ f · exp −M(τ )/60 ˜ ˜ ≤ 121d˜f /δ f · exp −M(τ )/60 . This completes the proof for r = r1/6 . Monotonicity extends the result to any r ∈ 0, r1/6 . Lemma 41 There exist (C, P, f , γ )-dependent constants τ ∗ ∈ N and c(ii) ∈ (0, ∞) such that, for any (ii) (i) integer τ ≥ τ ∗ , there is an event Hτ ⊆ Gτ with (i) (ii) P Hτ \ Hτ ˜ ≤ c(ii) · exp −M(τ )1/3 /60 (27) (i) (ii) such that, on Hτ ∩ Hτ , ∀s, m, ℓ, k ∈ N with ℓ < m and k ≤ d˜f , for any set of classifiers H with ⋆ ⊆ H, if either k = 1, or s ≥ τ and H ⊆ B( f , r Vℓ (1−γ )/6 ), then ˆ (k) ˆ (k) ˆ (k) ∆s (Xm ,W2 , H) < γ =⇒ Γs (Xm , − f (Xm ),W2 , H) < Γs (Xm , f (Xm ),W2 , H) . (i) (ii) In particular, for δ ∈ (0, 1) and τ ≥ max{τ ((1 − γ )/6; δ ), τ ∗ }, on Hτ (δ ) ∩ Hτ ∩ Hτ , this is true for H = Vℓ⋆ for every k, ℓ, m, s ∈ N satisfying τ ≤ ℓ < m, τ ≤ s, and k ≤ d˜f . ˜ Proof Let τ ∗ = (6/(1− γ ))· 2/δ f 1/3 , and consider any τ , k, ℓ, m, s, H as described above. If k = 1, (i) ⋆ the result clearly holds. In particular, Lemma 35 implies that on Hτ , H[(Xm , f (Xm ))] ⊇ Vm = ∅, so that some h ∈ H has h(Xm ) = f (Xm ), and therefore ˆ (1) Γs (Xm , − f (Xm ),W2 , H) = ½ {h(Xm )} (− f (Xm )) = 0, h∈H ˆ (1) ˆ (1) / and since ∆s (Xm ,W2 , H) = ½DIS(H) (Xm ), if ∆s (Xm ,W2 , H) < γ , then since γ < 1 we have Xm ∈ DIS(H), so that ˆ (1) Γs (Xm , f (Xm ),W2 , H) = ½ 1544 {h(Xm )} ( f (Xm )) h∈H = 1. ACTIVIZED L EARNING (i) (i) Otherwise, suppose 2 ≤ k ≤ d˜f . Note that on Hτ ∩ Gτ , ∀m ∈ N, and any H with Vℓ⋆ ⊆ H ⊆ B( f , r(1−γ )/6 ) for some ℓ ∈ N, ˆ (k) Γs (Xm , − f (Xm ),W2 , H) = (k) Ms (H) i=1 ½S¯k−1 (H[(Xm , f (Xm ))]) Si(k) ½S k−1 (H) Si(k) s3 1 ≤ (k) Si :i≤ s3 k−1 ∩ ∂H f (k) Si :i≤ s3 k−1 ∩ ∂H f = (k) Si :i≤ s3 i=1 s3 1 k−1 ∩ ∂C f s3 3 (k) 2Ms (B( f , r(1−γ )/6 )) i=1 ½S¯k−1 (Vm ) Si(k) ½S k−1 (B( f ,r(1−γ )/6 )) Si(k) ⋆ (monotonicity) ½∂ k−1 f Si(k) ½S k−1 (B( f ,r(1−γ )/6 )) Si(k) ¯ ⋆ (monotonicity) ½∂C f Si(k) ½S k−1 (B( f ,r(1−γ )/6 )) Si(k) ¯ k−1 (Lemma 35) i=1 s3 1 ≤ ≤ s3 1 Vm i=1 ½∂C f Si(k) ½S k−1 (B( f ,r(1−γ )/6 )) Si(k) . ¯ k−1 (Lemma 40) (k) ˆ For brevity, let Γ denote this last quantity, and let Mks = Ms B f , r(1−γ )/6 . By Hoeffding’s inequality, we have ¯ k−1 ˆ P (2/3)Γ > P k−1 ∂C f S k−1 B f , r(1−γ )/6 −1/3 + Mks 1/3 ≤ exp −2Mks Mks . Thus, by Lemmas 36, 39, and 40, P (i) (i) ˆ (k) ˜ (2/3)Γs (Xm , − f (Xm ),W2 , H) > q r(1−γ )/6 + M(s)−1/3 ∩ Hτ ∩ Gτ (i) ≤P ¯ k−1 ˆ (2/3)Γ > P k−1 ∂C f S k−1 B f , r(1−γ )/6 ˜ + M(s)−1/3 ∩ Hτ ≤P ¯ k−1 ˆ (2/3)Γ > P k−1 ∂C f S k−1 B f , r(1−γ )/6 + Mks ¯ k−1 ˆ = E P (2/3)Γ > P k−1 ∂C f S k−1 B f , r(1−γ )/6 1/3 ≤ E exp −2Mks −1/3 ˜ ∩ {Mks ≥ M(s)} −1/3 + Mks Mks ½[M(s),∞) (Mks ) ˜ ˜ ½[M(s),∞) (Mks ) ≤ exp −2M(s)1/3 . ˜ (i) (ii) (ii) (ii) ˜ Thus, there is an event Hτ (k, s) with P Hτ ∩ Gτ \ Hτ (k, s) ≤ exp −2M(s)1/3 such that ˆ (k) ˜ Γs (Xm , − f (Xm ),W2 , H) ≤ (3/2) q r(1−γ )/6 + M(s)−1/3 holds for these particular values of k and s. 1545 H ANNEKE (ii) (i) To extend to the full range of values, we simply take Hτ = Gτ ∩ ˜ ˜ τ ≥ (2/δ f )1/3 , we have M(τ ) ≥ 1, so a union bound implies (i) (i) (ii) P Hτ ∩ Gτ \ Hτ ≤ s≥τ s≥τ k≤d˜f (ii) Hτ (k, s). Since ˜ d˜f · exp −2M(s)1/3 ∞ ˜ ≤ d˜f · exp −2M(τ )1/3 + τ ˜ exp −2M(x)1/3 dx ˜ −1/3 · exp −2M(τ )1/3 ≤ 2d˜f δ −1/3 · exp −2M(τ )1/3 . ˜ ˜ ˜ = d˜f 1 + 2−2/3 δ f f Then Lemma 40 and a union bound imply (i) (ii) P Hτ \ Hτ ˜ ˜ −1/3 · exp −2M(τ )1/3 + 121d˜f δ −1 · exp −M(τ )/60 ˜ ˜ ≤ 2d˜f δ f f ˜ ˜ ≤ 123d˜f δ f−1 · exp −M(τ )1/3 /60 . (i) (ii) On Hτ ∩ Hτ , every such s, m, ℓ, k and H satisfy ˆ (k) ˜ Γs (Xm , − f (Xm ),W2 , H) ≤ (3/2) q(r(1−γ )/6 ) + M(s)−1/3 < (3/2) ((1 − γ )/6 + (1 − γ )/6) = (1 − γ )/2, (28) where the second inequality follows by definition of r(1−γ )/6 and s ≥ τ ≥ τ ∗ . ˆ (k) If ∆s (Xm ,W2 , H) < γ , then 1 ˆ (k) 1 − γ < 1 − ∆s (Xm ,W2 , H) = s3 (k) Ms (H) i=1 ½S k−1 (H) Si(k) ½S¯k (H) Si(k) ∪ {Xm } . Finally, noting that we always have ½S¯k (H) Si(k) ∪ {Xm } ≤ ½S¯k−1 (H[(Xm , f (Xm ))]) Si(k) + ½S¯k−1 (H[(Xm ,− f (Xm ))]) Si(k) , (i) (ii) (k) ˆ we have that, on the event Hτ ∩ Hτ , if ∆s (Xm ,W2 , H) < γ , then ˆ (k) Γs (Xm , − f (Xm ),W2 , H) < (1 − γ )/2 = −(1 − γ )/2 + (1 − γ ) < −(1 − γ )/2 + ≤ −(1 − γ )/2 + s3 (k) Ms (H) i=1 1 s3 (k) Ms (H) i=1 1 s3 ½S k−1 (H) Si(k) ½S¯k (H) Si(k) ∪ {Xm } by (29) ½S k−1 (H) Si(k) ½S¯k−1 (H[(Xm , f (Xm ))]) Si(k) ½ ½ (k) (k) ¯ S k−1 (H) Si S k−1 (H[(Xm ,− f (Xm ))]) Si (k) Ms (H) i=1 ˆ (k) ˆ (k) −(1 − γ )/2 + Γs (Xm , − f (Xm ),W2 , H) + Γs (Xm , f (Xm ),W2 , H) + = 1 by (28) (k) ˆ < Γs (Xm , f (Xm ),W2 , H) . by (28) 1546 (29) ACTIVIZED L EARNING The final claim in the lemma statement is then implied by Lemma 29, since we have Vℓ⋆ ⊆ Vτ⋆ ⊆ B ( f , φ (τ ; δ )) ⊆ B f , r(1−γ )/6 on Hτ (δ ). For any k, ℓ, m ∈ N, and any x ∈ X , define (k) ˆ px (k, ℓ, m) = ∆m (x,W2 ,Vℓ⋆ ) ˆ px (k, ℓ) = P k−1 S ∈ X k−1 : S ∪ {x} ∈ S k (Vℓ⋆ ) S k−1 (Vℓ⋆ ) . Lemma 42 For any ζ ∈ (0, 1), there is a (C, P, f , ζ )-dependent constant c(iii) (ζ ) ∈ (0, ∞) such (iii) that, for any τ ∈ N, there is an event Hτ (ζ ) with (i) (iii) ˜ P Hτ \ Hτ (ζ ) ≤ c(iii) (ζ ) · exp −ζ 2 M(τ ) (iii) (i) such that on Hτ ∩ Hτ (ζ ), ∀k, ℓ, m ∈ N with τ ≤ ℓ ≤ m and k ≤ d˜f , for any x ∈ X , ˜ P (x : |px (k, ℓ) − px (k, ℓ, m)| > ζ ) ≤ exp −ζ 2 M(m) . ˆ Proof Fix any k, ℓ, m ∈ N with τ ≤ ℓ ≤ m and k ≤ d˜f . Recall our convention that X 0 = {∅} and P 0 X 0 = 1; thus, if k = 1, px (k, ℓ, m) = ½DIS(V ⋆ ) (x) = ½S 1 (V ⋆ ) (x) = px (k, ℓ), so the result clearly ˆ ℓ ℓ holds for k = 1. (k) For the remaining case, suppose 2 ≤ k ≤ d˜f . To simplify notation, let m = Mm (Vℓ⋆ ), X = Xℓ+1 , ˜ px = px (k, ℓ) and px = px (k, ℓ, m). Consider the event ˆ ˆ ˜ ˆ H (iii) (k, ℓ, m, ζ ) = P (x : |px − px | > ζ ) ≤ exp −ζ 2 M(m) . We have (i) P Hτ \ H (iii) (k, ℓ, m, ζ ) Vℓ⋆ (30) ≤P ˜ m ≥ M(m) \ H (iii) (k, ℓ, m, ζ ) Vℓ⋆ ˜ =P 2 ˜ ˜ ˆ ˜ ˜ m ≥ M(m) ∩ P esm|pX − pX | > esmζ W2 ,Vℓ⋆ > e−ζ M(m) ˜ (by Lemma 39) Vℓ⋆ , (31) for any value s > 0. Proceeding as in Chernoff’s bounding technique, by Markov’s inequality (31) is at most P ≤P =E ˜ ˜ ˆ ˜ m ≥ M(m) ∩ e−smζ E esm|pX − pX | W2 ,Vℓ⋆ > e−ζ ˜ 2 M(m) ˜ Vℓ⋆ ˜ ˆ ˜ ˆ ˜ ˜ m ≥ M(m) ∩ e−smζ E esm(pX − pX ) + esm( pX −pX ) W2 ,Vℓ⋆ > e−ζ ˜ ˜ ˜ ˆ ˜ ˆ ˜ ½[M(m),∞) (m) P e−smζ E esm(pX − pX ) + esm( pX −pX ) W2 ,Vℓ⋆ > e−ζ ˜ 1547 2 M(m) ˜ 2 M(m) ˜ Vℓ⋆ m,Vℓ⋆ ˜ Vℓ⋆ H ANNEKE By Markov’s inequality, this is at most E ˜ ½[M(m),∞) (m) eζ ˜ 2 M(m) ˜ ˜ ˆ ˜ ˆ ˜ ˜ E e−smζ E esm(pX − pX ) + esm( pX −pX ) W2 ,Vℓ⋆ m,Vℓ⋆ Vℓ⋆ =E ˜ ½[M(m),∞) (m) eζ ˜ 2 M(m) ˜ ˜ ˆ ˜ ˆ ˜ ˜ e−smζ E esm(pX − pX ) + esm( pX −pX ) m,Vℓ⋆ Vℓ⋆ =E ˜ ½[M(m),∞) (m) eζ ˜ 2 M(m) ˜ ˜ ˆ ˜ ˆ ˜ ˜ ˜ e−smζ E E esm(pX − pX ) + esm( pX −pX ) X, m,Vℓ⋆ m,Vℓ⋆ Vℓ⋆ . (32) ∞ ˜ The conditional distribution of m pX given (X, m,Vℓ⋆ ) is Binomial (m, pX ), so letting B j (pX ) j=1 ˜ˆ ˜ denote a sequence of random variables, conditionally independent given (X, m,Vℓ⋆ ), with the condi˜ tional distribution of each B j (pX ) being Bernoulli(pX ) given (X, m,Vℓ⋆ ), we have ˜ ˜ ˆ ˜ ˆ E esm(pX − pX ) + esm( pX −pX ) X, m,Vℓ⋆ ˜ ˜ ˆ ˜ ˆ ˜ = E esm(pX − pX ) X, m,Vℓ⋆ + E esm( pX −pX ) X, m,Vℓ⋆ ˜ m ˜ =E ˜ ∏ es(pX −Bi (pX )) X, m,Vℓ⋆ + E i=1 m ˜ = E es(pX −B1 (pX )) X, m,Vℓ⋆ ˜ m ˜ ˜ ∏ es(B (p )−p ) X, m,Vℓ⋆ i X X i=1 + E es(B1 (pX )−pX ) X, m,Vℓ⋆ ˜ m ˜ . (33) 2 It is known that for B ∼ Bernoulli(p), E es(B−p) and E es(p−B) are at most es /8 (see, e.g., Lemma ˜ 2 8.1 of Devroye, Gy¨ rfi, and Lugosi, 1996). Thus, taking s = 4ζ , (33) is at most 2e2mζ , and (32) is o at most E ˜ ½[M(m),∞) (m) 2eζ ˜ 2 M(m) ˜ ˜ ˜ e−4mζ e2mζ Vℓ⋆ = E 2 2 ˜ ½[M(m),∞) (m) 2eζ ˜ 2 M(m) ˜ ˜ e−2mζ Vℓ⋆ 2 ˜ ≤ 2 exp −ζ 2 M(m) . Since this bound holds for (30), the law of total probability implies (i) (i) P Hτ \ H (iii) (k, ℓ, m, ζ ) = E P Hτ \ H (iii) (k, ℓ, m, ζ ) Vℓ⋆ 1548 ˜ ≤ 2 · exp −ζ 2 M(m) . ACTIVIZED L EARNING d˜ (iii) f Defining Hτ (ζ ) = ℓ≥τ m≥ℓ k=2 H (iii) (k, ℓ, m, ζ ), we have the required property for the claimed ranges of k, ℓ and m, and a union bound implies (iii) (i) P Hτ \ Hτ (ζ ) ≤ ≤ 2d˜f · = 2d˜f · ℓ≥τ ℓ≥τ ℓ≥τ m≥ℓ ˜ 2d˜f · exp −ζ 2 M(m) ˜ exp −ζ 2 M(ℓ) + ∞ ℓ3 ˜ exp −xζ 2 δ f /2 dx ˜ ˜ 1 + 2ζ −2 δ f−1 · exp −ζ 2 M(ℓ) ˜ ˜ ≤ 2d˜f · 1 + 2ζ −2 δ f−1 · exp −ζ 2 M(τ ) + ˜ = 2d˜f · 1 + 2ζ −2 δ f−1 2 ∞ τ3 ˜ exp −xζ 2 δ f /2 dx ˜ · exp −ζ 2 M(τ ) ˜ ˜ ≤ 18d˜f ζ −4 δ f−2 · exp −ζ 2 M(τ ) . For k, ℓ, m ∈ N and ζ ∈ (0, 1), define ˆ pζ (k, ℓ, m) = P (x : px (k, ℓ, m) ≥ ζ ) . ¯ (34) √ (i) Lemma 43 For any α , ζ , δ ∈ (0, 1), β ∈ 0, 1 − α , and integer τ ≥ τ (β ; δ ), on Hτ (δ ) ∩ Hτ ∩ (iii) Hτ (β ζ ), for any k, ℓ, ℓ′ , m ∈ N with τ ≤ ℓ ≤ ℓ′ ≤ m and k ≤ d˜f , ˜ pζ (k, ℓ′ , m) ≤ P (x : px (k, ℓ) ≥ αζ ) + exp −β 2 ζ 2 M(m) . ¯ (35) √ Proof Fix any α , ζ , δ ∈ (0, 1), β ∈ 0, 1 − α , τ , k, ℓ, ℓ′ , m ∈ N with τ (β ; δ ) ≤ τ ≤ ℓ ≤ ℓ′ ≤ m and k ≤ d˜f . If k = 1, the result clearly holds. In particular, we have pζ (1, ℓ′ , m) = P (DIS (Vℓ⋆ )) ≤ P (DIS (Vℓ⋆ )) = P (x : px (1, ℓ) ≥ αζ ) . ¯ ′ Otherwise, suppose 2 ≤ k ≤ d˜f . By a union bound, pζ (k, ℓ′ , m) = P x : px (k, ℓ′ , m) ≥ ζ ¯ ˆ √ √ ≤ P x : px (k, ℓ′ ) ≥ αζ + P x : px (k, ℓ′ ) − px (k, ℓ′ , m) > (1 − α )ζ . ˆ (36) Since √ ˆ P x : px (k, ℓ′ ) − px (k, ℓ′ , m) > (1 − α )ζ ≤ P x : px (k, ℓ′ ) − px (k, ℓ′ , m) > β ζ , ˆ (i) (iii) Lemma 42 implies that, on Hτ ∩ Hτ (β ζ ), √ ˜ ˆ P x : px (k, ℓ′ ) − px (k, ℓ′ , m) > (1 − α )ζ ≤ exp −β 2 ζ 2 M(m) . 1549 (37) H ANNEKE It remains only to examine the first term on the right side of (36). For this, if P k−1 S k−1 Vℓ⋆ = 0, ′ then the first term is 0 by our aforementioned convention, and thus (35) holds; otherwise, since ∀x ∈ X , S ∈ X k−1 : S ∪ {x} ∈ S k (Vℓ⋆ ) ⊆ S k−1 (Vℓ⋆ ) , ′ ′ we have = P x : P k−1 √ αζ = P x : P k−1 S ∈ X k−1 : S ∪ {x} ∈ S k (Vℓ⋆ ) S k−1 (Vℓ⋆ ) ≥ ′ ′ √ S ∈ X k−1 : S ∪ {x} ∈ S k (Vℓ⋆ ) ≥ αζ P k−1 S k−1 (Vℓ⋆ ) . ′ ′ P x : px (k, ℓ′ ) ≥ √ αζ (38) (i) By Lemma 35 and monotonicity, on Hτ ⊆ H ′ , (38) is at most √ k−1 P x : P k−1 S ∈ X k−1 : S ∪ {x} ∈ S k (Vℓ⋆ ) ≥ αζ P k−1 ∂C f ′ , and monotonicity implies this is at most P x : P k−1 S ∈ X k−1 : S ∪ {x} ∈ S k (Vℓ⋆ ) ≥ √ k−1 αζ P k−1 ∂C f . (39) (i) By Lemma 36, for τ ≥ τ (β ; δ ), on Hτ (δ ) ∩ Hτ , √ ¯ k−1 P k−1 ∂C f S k−1 (Vℓ⋆ ) ≤ q(φ (τ ; δ )) < β ≤ 1 − α , which implies k−1 k−1 P k−1 ∂C f ≥ P k−1 ∂C f ∩ S k−1 (Vℓ⋆ ) ¯ k−1 = 1 − P k−1 ∂C f S k−1 (Vℓ⋆ ) P k−1 S k−1 (Vℓ⋆ ) ≥ √ k−1 k−1 ⋆ αP S (Vℓ ) . (i) Altogether, for τ ≥ τ (β ; δ ), on Hτ (δ ) ∩ Hτ , (39) is at most P x : P k−1 S ∈ X k−1 : S∪{x} ∈ S k (Vℓ⋆ ) ≥ αζ P k−1 S k−1 (Vℓ⋆ ) = P (x : px (k, ℓ) ≥ αζ ), which, combined with (36) and (37), establishes (35). (iv) Lemma 44 There are events Hτ : τ ∈ N with (iv) P Hτ ≥ 1 − 3d˜f · exp {−2τ } s.t. for any ξ ∈ (0, γ /16], δ ∈ (0, 1), letting τ (iv) (ξ ; δ ) = max τ (4ξ /γ ; δ ), 4 ˜ δf ξ2 ln 4 ˜ δf ξ2 1/3 , (i) (iii) (iv) for any integer τ ≥ τ (iv) (ξ ; δ ), on Hτ (δ ) ∩ Hτ ∩ Hτ (ξ ) ∩ Hτ , ∀k ∈ 1, . . . , d˜f , ∀ℓ ∈ N with ℓ ≥ τ, ˆ (k) ˜ P x : px (k, ℓ) ≥ γ /2 + exp −γ 2 M(ℓ)/256 ≤ ∆ℓ (W1 ,W2 ,Vℓ⋆ ) ≤ P (x : px (k, ℓ) ≥ γ /8) + 4ℓ−1 . 1550 (40) (41) ACTIVIZED L EARNING Proof For any k, ℓ ∈ N, by Hoeffding’s inequality and the law of total probability, on an event G(iv) (k, ℓ) with P G(iv) (k, ℓ) ≥ 1 − 2 exp {−2ℓ}, we have ℓ3 pγ /4 (k, ℓ, ℓ) − ℓ ¯ (iv) Define the event Hτ = (iv) 1 − P Hτ −3 i=1 ˆ ½[γ /4,∞) ∆(k) (wi ,W2 ,Vℓ⋆ ) ≤ ℓ−1 . ℓ d˜f (iv) (k, ℓ). k=1 G ℓ≥τ ≤ 2d˜f · ℓ≥τ (42) By a union bound, we have exp {−2ℓ} ≤ 2d˜f · exp {−2τ } + ∞ τ exp {−2x} dx = 3d˜f · exp {−2τ } . Now fix any ℓ ≥ τ and k ∈ 1, . . . , d˜f . By a union bound, P (x : px (k, ℓ) ≥ γ /2) ≤ P (x : px (k, ℓ, ℓ) ≥ γ /4) + P (x : |px (k, ℓ) − px (k, ℓ, ℓ)| > γ /4) . ˆ ˆ (i) (43) (iii) By Lemma 42, on Hτ ∩ Hτ (ξ ), ˜ P (x : |px (k, ℓ) − px (k, ℓ, ℓ)| > γ /4) ≤ P (x : |px (k, ℓ) − px (k, ℓ, ℓ)| > ξ ) ≤ exp −ξ 2 M(ℓ) . (44) ˆ ˆ (iv) Also, on Hτ , (42) implies P (x : px (k, ℓ, ℓ) ≥ γ /4) = pγ /4 (k, ℓ, ℓ) ˆ ¯ ℓ3 ≤ℓ −1 +ℓ −3 ˆ ½[γ /4,∞) ∆(k) (wi ,W2 ,Vℓ⋆ ) ℓ i=1 ˆ (k) = ∆ℓ (W1 ,W2 ,Vℓ⋆ ) − ℓ−1 . (45) Combining (43) with (44) and (45) yields (k) ˆ ˜ P (x : px (k, ℓ) ≥ γ /2) ≤ ∆ℓ (W1 ,W2 ,Vℓ⋆ ) − ℓ−1 + exp −ξ 2 M(ℓ) . (46) ˜ ˜ For τ ≥ τ (iv) (ξ ; δ ), exp −ξ 2 M(ℓ) − ℓ−1 ≤ − exp −γ 2 M(ℓ)/256 , so that (46) implies the first inequality of the lemma: namely (40). (iv) For the second inequality (i.e., (41)), on Hτ , (42) implies we have ˆ (k) ∆ℓ (W1 ,W2 ,Vℓ⋆ ) ≤ pγ /4 (k, ℓ, ℓ) + 3ℓ−1 . ¯ (47) √ Also, by Lemma 43 (with α = 1/2, ζ = γ /4, β = ξ /ζ < 1 − α ), for τ ≥ τ (iv) (ξ ; δ ), on Hτ (δ ) ∩ (iii) (i) Hτ ∩ Hτ (ξ ), ˜ pγ /4 (k, ℓ, ℓ) ≤ P (x : px (k, ℓ) ≥ γ /8) + exp −ξ 2 M(ℓ) . ¯ (48) Thus, combining (47) with (48) yields ˆ (k) ˜ ∆ℓ (W1 ,W2 ,Vℓ⋆ ) ≤ P (x : px (k, ℓ) ≥ γ /8) + 3ℓ−1 + exp −ξ 2 M(ℓ) . 1551 H ANNEKE ˜ For τ ≥ τ (iv) (ξ ; δ ), we have exp −ξ 2 M(ℓ) ≤ ℓ−1 , which establishes (41). For n ∈ N and k ∈ {1, . . . , d + 1}, define the set (k) (k) ˆ Un = mn + 1, . . . , mn + n/ 6 · 2k ∆mn (W1 ,W2 ,V ) , (k) where mn = ⌊n/3⌋; Un represents the set of indices processed in the inner loop of Meta-Algorithm 1 for the specified value of k. ˆ ˆ Lemma 45 There are ( f , C, P, γ )-dependent constants c1 , c2 ∈ (0, ∞) such that, for any ε ∈ (0, 1) ˆ and integer n ≥ c1 ln(c2 /ε ), on an event Hn (ε ) with ˆ ˆ ˆ P(Hn (ε )) ≥ 1 − (3/4)ε , (49) ⋆ we have, for V = Vmn , (k) (k) ˆ m ∈ Un : ∆m (Xm ,W2 ,V ) ≥ γ ∀k ∈ 1, . . . , d˜f , ≤ n/ 3 · 2k , (50) (γ /8) ˆ (d ) ∆mnf (W1 ,W2 ,V ) ≤ ∆n (ε ) + 4m−1 , n ˜ (d˜f ) and ∀m ∈ Un (51) , ˜ ˜ ˜ ˆ (d ) ˆ (d ) ˆ (d ) ∆m f (Xm ,W2 ,V ) < γ ⇒ Γm f (Xm , − f (Xm ),W2 ,V ) < Γm f (Xm , f (Xm ),W2 ,V ). (52) Proof Suppose n ≥ c1 ln(c2 /ε ), where ˆ ˆ ˜ c1 = max ˆ 2d f +12 24 24 , , 3τ ∗ , ˜ f γ 2 r(1/16) r(1−γ )/6 δ and c2 = max 4 c(i) + c(ii) + c(iii) (γ /16) + 6d˜f , 4 ˆ 4e r(1/16) d ,4 4e d r(1−γ )/6 . In particular, we have chosen c1 and c2 large enough so that ˆ ˆ mn ≥ max τ (1/16; ε /2), τ (iv) (γ /16; ε /2), τ ((1 − γ )/6; ε /2), τ ∗ . We begin with (50). By Lemmas 43 and 44, on the event (iii) (i) (iv) ˆ (1) Hn (ε ) = Hmn (ε /2) ∩ Hmn ∩ Hmn (γ /16) ∩ Hmn , (k) ∀m ∈ Un , ∀k ∈ 1, . . . , d˜f , ˜ pγ (k, mn , m) ≤ P (x : px (k, mn ) ≥ γ /2) + exp −γ 2 M(m)/256 ¯ (k) ˆ ˜ ≤ P (x : px (k, mn ) ≥ γ /2) + exp −γ 2 M(mn )/256 ≤ ∆mn (W1 ,W2 ,V ) . 1552 (53) ACTIVIZED L EARNING Recall that (k) (k) ˆ is a sample of size n/(6 · 2k ∆mn (W1 ,W2 ,V )) , conditionally i.i.d. Xm : m ∈ Un (1) ˆ (given (W1 ,W2 ,V )) with conditional distributions P. Thus, ∀k ∈ 1, . . . , d˜f , on Hn (ε ), P ≤P (k) (k) ˆ m ∈ Un : ∆m (Xm ,W2 ,V ) ≥ γ > n/ 3 · 2k (k) (k) ˆ (k) m ∈ Un : ∆m (Xm ,W2 ,V ) ≥ γ (k) > 2 Un (k) (k) ˆ ≤ P B |Un |, ∆mn (W1 ,W2 ,V ) > 2 Un W1 ,W2 ,V ˆ (k) ∆mn (W1 ,W2 ,V ) W1 ,W2 ,V ˆ (k) ∆mn (W1 ,W2 ,V ) W1 ,W2 ,V , (54) where this last inequality follows from (53), and B(u, p) ∼ Binomial(u, p) is independent from W1 ,W2 ,V (for any fixed u and p). By a Chernoff bound, (54) is at most (k) ˆ exp − n/ 6 · 2k ∆mn (W1 ,W2 ,V ) ˆ (k) ∆mn (W1 ,W2 ,V )/3 ≤ exp 1 − n/ 18 · 2k . ˆ (2) By the law of total probability and a union bound, there exists an event Hn with ˆ (2) ≤ d˜f · exp 1 − n/ 18 · 2d˜f ˆ (1) P Hn (ε ) \ Hn ˆ (1) ˆ (2) such that, on Hn (ε ) ∩ Hn , (50) holds. ˆ (1) Next, by Lemma 44, on Hn (ε ), ˜ ˆ (d ) ∆mnf (W1 ,W2 ,V ) ≤ P x : px d˜f , mn ≥ γ /8 + 4m−1 , n (γ /8) (1) ˆ and by Lemma 38, on Hn (ε ), this is at most ∆n (ε ) + 4m−1 , which establishes (51). n (˜ (1) (ii) ˆ n (ε ) ∩ Hmn , ∀m ∈ Und f ) , (52) holds. Finally, Lemma 41 implies that on H Thus, defining (ii) ˆ (1) ˆ (2) ˆ Hn (ε ) = Hn (ε ) ∩ Hn ∩ Hmn , it remains only to establish (49). By a union bound, we have (i) (i) ˆ 1 − P Hn ≤ (1 − P (Hmn (ε /2))) + 1 − P Hmn (i) (iii) (ii) + P Hmn \ Hmn (iv) + P Hmn \ Hmn (γ /16) + 1 − P Hmn ˆ (1) ˆ (2) + P Hn (ε ) \ Hn . ˜ ˜ ≤ ε /2 + c(i) · exp −M(mn )/4 + c(ii) · exp −M(mn )1/3 /60 ˜ + c(iii) (γ /16) · exp −M(mn )γ 2 /256 + 3d˜f · exp {−2mn } ˜ + d˜f · exp 1 − n/ 18 · 2d f ˜ ˜ ≤ ε /2 + c(i) + c(ii) + c(iii) (γ /16) + 6d˜f · exp −nδ f γ 2 2−d f −12 . We have chosen n large enough so that (55) is at most (3/4)ε , which establishes (49). The following result is a slightly stronger version of Theorem 6. 1553 (55) H ANNEKE Lemma 46 For any passive learning algorithm A p , if A p achieves a label complexity Λ p with ∞ > Λ p (ε , f , P) = ω (log(1/ε )), then Meta-Algorithm 1, with A p as its argument, achieves a label complexity Λa such that Λa (3ε , f , P) = o(Λ p (ε , f , P)). Proof Suppose A p achieves label complexity Λ p with ∞ > Λ p (ε , f , P) = ω (log(1/ε )). Let ε ∈ (γ /8) ˜ (0, 1), define L(n; ε ) = n/ 6 · 2d f ∆n (ε ) + 4m−1 n max {n ∈ N : L(n; ε ) < m} (for any m ∈ (0, ∞)). Define c1 = max c1 , 2 · 63 (d + 1)d˜f ln(e(d + 1)) ˆ (for any n ∈ N), and let L−1 (m; ε ) = and c2 = max {c2 , 4e(d + 1)} , ˆ and suppose n ≥ max c1 ln(c2 /ε ), 1 + L−1 (Λ p (ε , f , P); ε ) . Consider running Meta-Algorithm 1 with A p and n as inputs, while f is the target function and P is the data distribution. ˆ Letting hn denote the classifier returned from Meta-Algorithm 1, Lemma 34 implies that on an ˆ ˆ event En with P(En ) ≥ 1 − e(d + 1) · exp −⌊n/3⌋/(72d˜f (d + 1) ln(e(d + 1))) ≥ 1 − ε /4, we have ˆ er(hn ) ≤ 2 er A p Ld˜f . ˆ ˆ ˆ ˆ By a union bound, the event Gn (ε ) = En ∩ Hn (ε ) has P Gn (ε ) ≥ 1 − ε . Thus, ˆ E er hn ≤E ˆ ½Gn (ε ) ½ |Ld˜f | ≥ Λ p (ε , f , P) er hn ˆ ˆ + P Gn (ε ) ∩ |Ld˜f | < Λ p (ε , f , P) ≤E ˆ + P Gn (ε )c ½Gn (ε ) ½ |Ld˜f | ≥ Λ p (ε , f , P) 2 er A p Ld˜f ˆ ˆ + P Gn (ε ) ∩ |Ld˜f | < Λ p (ε , f , P) + ε. (56) ˆ On Gn (ε ), (51) of Lemma 45 implies |Ld˜f | ≥ L(n; ε ), and we chose n large enough so that L(n; ε ) ≥ Λ p (ε , f , P). Thus, the second term in (56) is zero, and we have ˆ E er hn ≤ 2·E ½Gn (ε ) ½ |Ld˜f | ≥ Λ p (ε , f , P) er A p Ld˜f ˆ = 2·E E ½Gn (ε ) er A p Ld˜f ˆ |Ld˜f | +ε ½ |Ld˜f | ≥ Λ p (ε , f , P) + ε . (d˜f ) Note that for any ℓ with P(|Ld˜f | = ℓ) > 0, the conditional distribution of Xm : m ∈ Un (57) given |Ld˜f | = ℓ is simply the product P ℓ (i.e., conditionally i.i.d.), which is the same as the distribution ˆ of {X1 , X2 , . . . , Xℓ }. Furthermore, on Gn (ε ), (50) implies that the t < ⌊2n/3⌋ condition is always satisfied in Step 6 of Meta-Algorithm 1 while k ≤ d˜f , and (52) implies that the inferred labels from Step 8 for k = d˜f are all correct. Therefore, for any such ℓ with ℓ ≥ Λ p (ε , f , P), we have E ½Gn (ε ) er A p Ld˜f ˆ |Ld˜f | = ℓ 1554 ≤ E [er (A p (Zℓ ))] ≤ ε . ACTIVIZED L EARNING In particular, this means (57) is at most 3ε . This implies that Meta-Algorithm 1, with A p as its argument, achieves a label complexity Λa such that Λa (3ε , f , P) ≤ max c1 ln(c2 /ε ), 1 + L−1 (Λ p (ε , f , P); ε ) . Since Λ p (ε , f , P) = ω (log(1/ε )) ⇒ c1 ln(c2 /ε ) = o (Λ p (ε , f , P)), it remains only to show that L−1 (Λ p (ε , f , P); ε ) = o (Λ p (ε , f , P)). Note that ∀ε ∈ (0, 1), L(1; ε ) = 0 and L(n; ε ) is diverging in n. Furthermore, by Lemma 38, we know that for any N-valued N(ε ) = ω (log(1/ε )), we have (γ /8) ∆N(ε ) (ε ) = o(1), which implies L(N(ε ); ε ) = ω (N(ε )). Thus, since Λ p (ε , f , P) = ω (log(1/ε )), Lemma 31 implies L−1 (Λ p (ε , f , P); ε ) = o (Λ p (ε , f , P)), as desired. This establishes the result for an arbitrary γ ∈ (0, 1). To specialize to the specific procedure stated as Meta-Algorithm 1, we simply take γ = 1/2. Proof [Theorem 6] Theorem 6 now follows immediately from Lemma 46. Specifically, we have proven Lemma 46 for an arbitrary distribution P on X , an arbitrary f ∈ cl(C), and an arbitrary passive algorithm A p . Therefore, it will certainly hold for every P and f ∈ C, and since every ( f , P) ∈ Nontrivial(Λ p ) has ∞ > Λ p (ε , f , P) = ω (log(1/ε )), the implication that Meta-Algorithm 1 activizes every passive algorithm A p for C follows. Careful examination of the proofs above reveals that the “3” in Lemma 46 can be set to any arbitrary constant strictly larger than 1, by an appropriate modification of the “7/12” threshold ˆ in ActiveSelect. In fact, if we were to replace Step 4 of ActiveSelect by instead selecting k = argmink max j=k mk j (where mk j = erQk j (hk ) when k < j), then we could even make this a certain (1 + o(1)) function of ε , at the expense of larger constant factors in Λa . Appendix C. The Label Complexity of Meta-Algorithm 2 As mentioned, Theorem 10 is essentially implied by the details of the proof of Theorem 16 in Appendix D below. Here we present a proof of Theorem 13, along with two useful related lemmas. The first, Lemma 47, lower bounds the expected number of label requests Meta-Algorithm 2 would make while processing a given number of random unlabeled examples. The second, Lemma 48, bounds the amount by which each label request is expected to reduce the probability mass in the region of disagreement. Although we will only use Lemma 48 in our proof of Theorem 13, Lemma 47 may be of independent interest, as it provides additional insights into the behavior of disagreement based methods, as related to the disagreement coefficient, and is included for this reason. Throughout, we fix an arbitrary class C, a target function f ∈ C, and a distribution P, and we ⋆ continue using the notational conventions of the proofs above, such as Vm = {h ∈ C : ∀i ≤ m, h(Xi ) = f (Xi )} (with V0⋆ = C). Additionally, for t ∈ N, define the random variable m M(t) = min m ∈ N : ℓ=1 ½DIS(Vℓ−1 ) (Xℓ ) = t , ⋆ which represents the index of the t th unlabeled example Meta-Algorithm 2 would request the label of (assuming it has not yet halted). The two aforementioned lemmas are formally stated as follows. 1555 H ANNEKE Lemma 47 For any r ∈ (0, 1) and ℓ ∈ N, E [P (DIS (Vℓ⋆ ∩ B ( f , r)))] ≥ (1 − r)ℓ P (DIS (B ( f , r))) , ⌈1/r⌉ P (DIS (B( f , r))) ½DIS(Vm−1 ∩B( f ,r)) (Xm ) ≥ E . ⋆ 2r and m=1 Lemma 48 For any r ∈ (0, 1) and n ∈ N, ⋆ E P DIS VM(n) ∩ B ( f , r) Note these results immediately imply that ⌈1/r⌉ E and ≥ P (DIS (B( f , r))) − nr. ½DIS(Vm−1 ) (Xm ) ≥ ⋆ m=1 ⋆ E P DIS VM(n) P (DIS (B( f , r))) 2r ≥ P (DIS (B( f , r))) − nr, which are then directly relevant to the expected number of label requests made by Meta-Algorithm 2 among the first m data points, and the probability Meta-Algorithm 2 requests the label of the next point, after already making n label requests, respectively. Before proving these lemmas, let us first mention their relevance to the disagreement coefficient analysis. Specifically, for any ε ∈ (0, r], we have ⌈1/ε ⌉ ⌈1/r⌉ P (DIS (B( f , r))) E ½DIS(Vm−1 ) (Xm ) ≥ E ½DIS(Vm−1 ) (Xm ) ≥ . ⋆ ⋆ 2r m=1 m=1 In particular, maximizing over r > ε , we have ⌈1/ε ⌉ E m=1 ½DIS(Vm−1 ) (Xm ) ≥ θ f (ε )/2. ⋆ Thus, the expected number of label requests among the first ⌈1/ε ⌉ unlabeled examples processed by Meta-Algorithm 2 is at least θ f (ε )/2 (assuming it does not halt first). Similarly, for any ε ∈ (0, r], for any n ≤ P(DIS(B( f , r)))/(2r), Lemma 48 implies ⋆ E P DIS VM(n) ≥ P (DIS (B( f , r))) /2 ≥ P (DIS (B( f , ε ))) /2. Maximizing over r > ε , we see that ⋆ n ≤ θ f (ε )/2 =⇒ E P DIS VM(n) ≥ P (DIS (B( f , ε ))) /2. In other words, for Meta-Algorithm 2 to arrive at a region of disagreement with expected probability mass less than P(DIS(B( f , ε )))/2 requires a budget n of at least θ f (ε )/2. 1556 ACTIVIZED L EARNING We now present proofs of Lemmas 47 and 48. ⋆ Proof [Lemma 47] Let Dm = DIS (Vm ∩ B( f , r)). Since ⌈1/r⌉ E m=1 ½Dm−1 (Xm ) = = ⌈1/r⌉ m=1 ⌈1/r⌉ ⋆ E P Xm ∈ Dm−1 Vm−1 E [P (Dm−1 )] , (58) m=1 we focus on lower bounding E [P (Dm )] for m ∈ N ∪ {0}. Note that for any x ∈ DIS(B( f , r)), there ⋆ exists some hx ∈ B( f , r) with hx (x) = f (x), and if this hx ∈ Vm , then x ∈ Dm as well. This means ⋆ ∀x, ½Dm (x) ≥ ½DIS(B( f ,r)) (x) · ½Vm (hx ) = ½DIS(B( f ,r)) (x) · ∏m ½DIS({hx , f })c (Xℓ ). Therefore, ℓ=1 E [P (Dm )] = P (Xm+1 ∈ Dm ) = E E ½Dm (Xm+1 ) Xm+1 m ≥E E ½DIS(B( f ,r)) (Xm+1 ) · ∏ ½DIS({hXm+1 , f })c (Xℓ ) Xm+1 ℓ=1 m =E ∏P hXm+1 (Xℓ ) = f (Xℓ ) Xm+1 ½DIS(B( f ,r)) (Xm+1 ) (59) ℓ=1 ≥ E (1 − r)m ½DIS(B( f ,r)) (Xm+1 ) = (1 − r)m P(DIS(B( f , r))), (60) where the equality in (59) is by conditional independence of the ½DIS({hXm+1 , f })c (Xℓ ) indicators, given Xm+1 , and the inequality in (60) is due to hXm+1 ∈ B( f , r). This indicates (58) is at least ⌈1/r⌉ m=1 (1 − r)m−1 P (DIS (B( f , r))) = 1 − (1 − r)⌈1/r⌉ ≥ 1− 1 e P (DIS (B( f , r))) r P (DIS (B( f , r))) P (DIS (B( f , r))) ≥ . r 2r ⋆ Proof [Lemma 48] For each m ∈ N ∪ {0}, let Dm = DIS (B( f , r) ∩Vm ). For convenience, let M(0) = 0. We prove the result by induction. We clearly have E P DM(0) = E [P (D0 )] = P(DIS(B( f , r))), which serves as our base case. Now fix any n ∈ N and take as the inductive hypothesis that E P DM(n−1) ≥ P(DIS(B( f , r))) − (n − 1)r. ⋆ As in the proof of Lemma 47, for any x ∈ DM(n−1) , there exists hx ∈ B( f , r) ∩VM(n−1) with hx (x) = ⋆ f (x); unlike the proof of Lemma 47, here hx is a random variable, determined by VM(n−1) . If hx is ⋆ ⋆ also in VM(n) , then x ∈ DM(n) as well. Thus, ∀x, ½DM(n) (x) ≥ ½DM(n−1) (x) · ½VM(n) (hx ) = ½DM(n−1) (x) · ½DIS({hx , f })c (XM(n) ), where this last equality is due to the fact that every m ∈ {M(n − 1) + 1, . . . , ⋆ M(n) − 1} has Xm ∈ DIS Vm−1 , so that in particular hx (Xm ) = f (Xm ). Therefore, letting X ∼ P be / 1557 H ANNEKE independent of the data Z, =E ½DM(n) (X) ≥ E ½DM(n−1) (X) · ½DIS({hX , f })c (XM(n) ) =E E P DM(n) ⋆ ½DM(n−1) (X) · P hX (XM(n) ) = f (XM(n) ) X,VM(n−1) . (61) ⋆ The conditional distribution of XM(n) given VM(n−1) is merely P but with support restricted to ⋆ DIS VM(n−1) ⋆ and renormalized to a probability measure: that is P · DIS VM(n−1) . Thus, ⋆ since any x ∈ DM(n−1) has DIS({hx , f }) ⊆ DIS VM(n−1) , we have ⋆ P hx (XM(n) ) = f (XM(n) ) VM(n−1) = P (DIS({hx , f })) ⋆ P DIS VM(n−1) ≤ r P DM(n−1) , ⋆ where the inequality follows from hx ∈ B( f , r) and DM(n−1) ⊆ DIS VM(n−1) . Therefore, (61) is at least E ½DM(n−1) (X)· 1 − r P(DM(n−1) ) = E P X ∈ DM(n−1) DM(n−1) · 1 − = E P DM(n−1) · 1 − r r P(DM(n−1) ) P(DM(n−1) ) = E P DM(n−1) − r. By the inductive hypothesis, this is at least P(DIS(B( f , r))) − nr. With Lemma 48 in hand, we are ready for the proof of Theorem 13. Proof [Theorem 13] Let C, f , P, and λ be as in the theorem statement. For m ∈ N, let λ −1 (m) = inf{ε > 0 : λ (ε ) ≤ m}, or 1 if this is not defined. We define A p as a randomized algorithm such that, for m ∈ N and L ∈ (X ×{−1, +1})m , A p (L) returns f with probability 1− λ −1 (|L|) and returns − f with probability λ −1 (|L|) (independent of the contents of L). Note that, for any integer m ≥ λ (ε ), E [er (A p (Zm ))] = λ −1 (m) ≤ λ −1 (λ (ε )) ≤ ε . Therefore, A p achieves some label complexity Λ p with Λ p (ε , f , P) = λ (ε ) for all ε > 0. If θ f λ (ε )−1 = ω (1), then monotonicity implies θ f λ (ε )−1 = O(1), and since every label complexity Λa is Ω(1), the result clearly holds. Otherwise, suppose θ f λ (ε )−1 = ω (1); in particular, this means ∃ε0 ∈ (0, 1/2) such that θ f λ (2ε0 )−1 ≥ 12. Fix any ε ∈ (0, ε0 ), let r > λ (2ε )−1 be such that P(DIS(B( f ,r))) ≥ θ f λ (2ε )−1 /2, and let n ∈ N satisfy n ≤ θ f λ (2ε )−1 /4. r ˆ Consider running Meta-Algorithm 2 with arguments A p and n, and let L denote the final value of the set L, and let m denote the value of m upon reaching Step 6. Note that any m < λ (2ε ) and ˇ m has er (A (L)) = λ −1 (m) ≥ inf{ε ′ > 0 : λ (ε ′ ) < λ (2ε )} ≥ 2ε . Therefore, L ∈ (X × {−1, +1}) p we have ˆ E er A p L ˆ ≥ 2ε P |L| < λ (2ε ) = 2ε P ˆ = 2ε P ∆ > n 6λ (2ε ) 1558 ˆ n/ 6∆ < λ (2ε ) ˆ = 2ε 1 − P ∆ ≤ n 6λ (2ε ) . (62) ACTIVIZED L EARNING Since n ≤ θ f λ (2ε )−1 /4 ≤ P(DIS(B( f , r)))/(2r) < λ (2ε )P(DIS(B( f , r)))/2, we have ˆ P ∆≤ n 6λ (2ε ) ˆ ≤ P ∆ < P(DIS(B( f , r)))/12 ⋆ ⋆ ˆ P (DIS (Vm )) < P(DIS(B( f , r)))/12 ∪ ∆ < P (DIS (Vm )) ˇ ˇ ≤P . (63) Since m ≤ M(⌈n/2⌉), monotonicity and a union bound imply this is at most ˇ ⋆ P P DIS VM(⌈n/2⌉) ⋆ ˆ < P(DIS(B( f , r)))/12 + P ∆ < P (DIS (Vm )) . ˇ (64) Markov’s inequality implies ⋆ P P DIS VM(⌈n/2⌉) < P(DIS(B( f , r)))/12 11 P(DIS(B( f , r))) 12 11 ⋆ ≤ P P(DIS(B( f , r))) − P DIS VM(⌈n/2⌉) ∩ B( f , r) > P(DIS(B( f , r))) 12 ⋆ = P P(DIS(B( f , r))) − P DIS VM(⌈n/2⌉) ≤ ⋆ E P(DIS(B( f , r))) − P DIS VM(⌈n/2⌉) ∩ B( f , r) 11 12 P(DIS(B( f , r))) ⋆ E P DIS VM(⌈n/2⌉) ∩ B( f , r) 12 = 1− 11 P(DIS(B( f , r))) Lemma 48 implies this is at most ⌈n/2⌉r 12 11 P(DIS(B( f ,r))) ≤ . 12 11 3/2 has ⌈a⌉ ≤ (3/2)a, and θ f λ (2ε )−1 ≥ 12 implies ≤ > 3 P(DIS(B( f ,r))) , 8 r so that 12 11 P(DIS(B( f ,r))) 4r ⋆ P P DIS VM(⌈n/2⌉) P(DIS(B( f ,r))) r 4r P(DIS(B( f ,r))) . Since any a ≥ P(DIS(B( f ,r))) ≥ 3/2, we have P(DIS(B( f ,r))) 4r 4r r P(DIS(B( f ,r))) ≤ 9 22 . Combining the above, we have < P(DIS(B( f , r)))/12 ≤ 9 . 22 (65) ˆ Examining the second term in (64), Hoeffding’s inequality and the definition of ∆ from (13) imply ⋆ ⋆ ⋆ ˆ ˆ P ∆ < P (DIS (Vm )) = E P ∆ < P (DIS (Vm )) Vm , m ˇ ˇ ˇ ˇ ˇ ≤ E e−8m ≤ e−8 < 1/11. (66) Combining (62), (63), (64), (65), and (66) implies ˆ E er A p L > 2ε 1 − 1 9 − 22 11 = ε. Thus, for any label complexity Λa achieved by running Meta-Algorithm 2 with A p as its argument, we must have Λa (ε , f , P) > θ f λ (2ε )−1 /4. Since this is true for all ε ∈ (0, ε0 ), this establishes the result. 1559 H ANNEKE Appendix D. The Label Complexity of Meta-Algorithm 3 As in Appendix B, we will assume C is a fixed VC class, P is some arbitrary distribution, and f ∈ cl(C) is an arbitrary fixed function. We continue using the notation introduced above: in k ˜ ¯k ¯ particular, S k (H) = S ∈ X k : H shatters S , S k (H) = X k \ S k (H), ∂H f = X k \ ∂H f , and δ f = ˜ d˜ −1 P d f −1 ∂Cf f . Also, as above, we will prove a more general result replacing the “1/2” in Steps 5, 9, and 12 of Meta-Algorithm 3 with an arbitrary value γ ∈ (0, 1); thus, the specific result for the stated algorithm will be obtained by taking γ = 1/2. ˆ For the estimators Pm in Meta-Algorithm 3, we take precisely the same definitions as given in ˆ (k) Appendix B.1 for the estimators in Meta-Algorithm 1. In particular, the quantities ∆m (x,W2 , H), (k) ˆ (k) ˆ ˆ (k) ∆m (W1 ,W2 , H), Γm (x, y,W2 , H), and Mm (H) are all defined as in Appendix B.1, and the Pm estimators are again defined as in (11), (12) and (13). Also, we sometimes refer to quantities defined above, such as pζ (k, ℓ, m) (defined in (34)), as ¯ (i) (ii) well as the various events from the lemmas of the previous appendix, such as Hτ (δ ), H ′ , Hτ , Hτ , (iii) (iv) (i) Hτ (ζ ), Hτ , and Gτ . D.1 Proof of Theorem 16 Throughout the proof, we will make reference to the sets Vm defined in Meta-Algorithm 3. Also let V (k) denote the final value of V obtained for the specified value of k in Meta-Algorithm 3. Both Vm and V (k) are implicitly functions of the budget, n, given to Meta-Algorithm 3. As above, we ⋆ continue to denote by Vm = {h ∈ C : ∀i ≤ m, h(Xm ) = f (Xm )}. One important fact we will use ⋆ ⋆ repeatedly below is that if Vm = Vm for some m, then since Lemma 35 implies that Vm = ∅ on H ′ , we must have that all of the previous y values were consistent with f , which means that ∀ℓ ≤ m, ˆ ⋆ . In particular, if V (k′ ) = V ⋆ for the largest m value obtained while k = k′ in Meta-Algorithm Vℓ = Vℓ m 3, then Vℓ = Vℓ⋆ for all ℓ obtained while k ≤ k′ in Meta-Algorithm 3. Additionally, define mn = ⌊n/24⌋, and note that the value m = ⌈n/6⌉ is obtained while k = 1 in ˜ Meta-Algorithm 3. We also define the following quantities, which we will show are typically equal ˆ to related quantities in Meta-Algorithm 3. Define m0 = 0, T0⋆ = ⌈2n/3⌉, and t0 = 0, and for each ˆ k ∈ {1, . . . , d + 1}, inductively define 1560 ACTIVIZED L EARNING ⋆ ˆ Tk⋆ = Tk−1 − tk−1 , ⋆ ⋆ ˆ Imk = ½[γ ,∞) ∆m Xm ,W2 ,Vm−1 , ∀m ∈ N, m ⋆ mk = min m ≥ mk−1 : ˇ ˆ Iℓk = ⌈Tk⋆ /4⌉ ∪ {max {k · 2n + 1, mk−1 }} , ˆ (k) ℓ=mk−1 +1 ˆ (k) ⋆ ˆ mk = mk + Tk⋆ / 3∆mk W1 ,W2 ,Vmk ˆ ˇ ˇ ˇ ˇ Uk = (mk−1 , mk ] ∩ N, ˆ ˇ ˆ Uk = (mk , mk ] ∩ N, ˇ ˆ ⋆ Cmk = ½[0,⌊3T ⋆ /4⌋) k Q⋆ k = ˆ m∈Uk m−1 ℓ=mk−1 +1 ˆ ⋆ ⋆ Imk ·Cmk , ˆ and tk = Q⋆ + k , ⋆ Iℓk ⋆ Imk . ˇ m∈Uk The meaning of these values can be understood in the context of Meta-Algorithm 3, under the ⋆ condition that Vm = Vm for values of m obtained for the respective value of k. Specifically, under ⋆ corresponds to T , t represents the final value t for round k, m represents the ˇk this condition, Tk k ˆk value of m upon reaching Step 9 in round k, while mk represents the value of m at the end of round k, ˆ ˇ ˆ Uk corresponds to the set of indices arrived at in Step 4 during round k, while Uk corresponds to the ⋆ indicates whether the label of X ˇ set of indices arrived at in Step 11 during round k, for m ∈ Uk , Imk m ˆk , I ⋆ · C⋆ indicates whether the label of Xm is requested. Finally Q⋆ is requested, while for m ∈ U mk mk k corresponds to the number of label requests in Step 13 during round k. In particular, note m1 ≥ mn . ˇ ˜ (i) Lemma 49 For any τ ∈ N, on the event H ′ ∩ Gτ , ∀k, ℓ, m ∈ N with k ≤ d˜f , ∀x ∈ X , for any sets H and H′ with Vℓ⋆ ⊆ H ⊆ H′ ⊆ B( f , r1/6 ), if either k = 1 or m ≥ τ , then ˆ (k) ˆ (k) ∆m (x,W2 , H) ≤ (3/2)∆m x,W2 , H′ . (i) In particular, for any δ ∈ (0, 1) and τ ≥ τ (1/6; δ ), on H ′ ∩ Hτ (δ ) ∩ Gτ , ∀k, ℓ, ℓ′ , m ∈ N with m ≥ τ , ˆ (k) ˆ (k) ℓ ≥ ℓ′ ≥ τ , and k ≤ d˜f , ∀x ∈ X , ∆m (x,W2 ,Vℓ⋆ ) ≤ (3/2)∆m x,W2 ,Vℓ⋆ . ′ Proof First note that ∀m ∈ N, ∀x ∈ X , ˆ (1) ˆ (1) ∆m (x,W2 , H) = ½DIS(H) (x) ≤ ½DIS(H′ ) (x) = ∆m x,W2 , H′ , (k) so the result holds for k = 1. Lemma 35, Lemma 40, and monotonicity of Mm (·) imply that on (i) H ′ ∩ Gτ , for any m ≥ τ and k ∈ 2, . . . , d˜f , m3 (k) Mm (H) ≥ i=1 (k) (k) ½∂C f Si(k) ≥ (2/3)Mm B( f , r1/6 ) ≥ (2/3)Mm H′ , k−1 1561 H ANNEKE so that ∀x ∈ X , m3 (k) ˆ (k) ∆m (x,W2 , H) = Mm (H)−1 i=1 m3 ≤ (k) Mm (H)−1 ≤ (k) (3/2)Mm i=1 H ½S k (H) Si(k) ∪ {x} ½S k (H′ ) Si(k) ∪ {x} ′ −1 m3 i=1 ˆm ½S k (H′ ) Si(k) ∪ {x} = (3/2)∆(k) x,W2 , H′ . The final claim follows from Lemma 29. ˆ Lemma 50 For any k ∈ {1, . . . , d + 1}, if n ≥ 3·4k−1 , then Tk⋆ ≥ 41−k (2n/3) and tk ≤ 3Tk⋆ /4 . Proof Recall T1⋆ = ⌈2n/3⌉ ≥ 2n/3. If n ≥ 2, we also have ⌊3T1⋆ /4⌋ ≥ ⌈T1⋆ /4⌉, so that (due to the ⋆ ˆ Cm1 factors) t1 ≤ ⌊3T1⋆ /4⌋. For the purpose of induction, suppose some k ∈ {2, . . . , d + 1} has n ≥ ⋆ ⋆ ⋆ ⋆ ˆ ˆ 3 · 4k−1 , Tk−1 ≥ 42−k (2n/3), and tk−1 ≤ ⌊3Tk−1 /4⌋. Then Tk⋆ = Tk−1 − tk−1 ≥ Tk−1 /4 ≥ 41−k (2n/3), ⋆ ˆ and since n ≥ 3 · 4k−1 , we also have ⌊3Tk⋆ /4⌋ ≥ ⌈Tk⋆ /4⌉, so that tk ≤ ⌊3Tk⋆ /4⌋ (again, due to the Cmk k−1 . factors). Thus, by induction, this holds for all k ∈ {1, . . . , d + 1} with n ≥ 3 · 4 The next lemma indicates that the “t < ⌊3Tk /4⌋” constraint in Step 12 is redundant for k ≤ d˜f . It ˆ is similar to (50) in Lemma 45, but is made only slightly more complicated by the fact that the ∆(k) estimate is calculated in Step 9 based on a set Vm different from the ones used to decide whether or not to request a label in Step 12. (i) (i) ˜ ˜ Lemma 51 There exist (C, P, f , γ )-dependent constants c1 , c2 ∈ [1, ∞) such that, for any δ ∈ (i) (i) (0, 1), and any integer n ≥ c1 ln c2 /δ , on an event ˜ ˜ (i) (i) (iv) (iii) ˜ (i) Hn (δ ) ⊆ Gmn ∩ Hmn (δ ) ∩ Hmn ∩ Hmn (γ /16) ∩ Hmn ˜ ˜ ˜ ˜ ˜ ˜ (i) ˆ with P Hn (δ ) ≥ 1 − 2δ , ∀k ∈ 1, . . . , d˜f , tk = mk ˆ m=mk−1 +1 ˆ ⋆ Imk ≤ 3Tk⋆ /4. Proof Define the constants (i) c1 = max ˜ (i) d˜ +6 192d 3·4 f ˜ r(3/32) , δ f γ 2 8e (i) , c2 = max ˜ r(3/32) ˜ , c(i) + c(iii) (γ /16) + 125d˜f δ f−1 (i) and let n(i) (δ ) = c1 ln c2 /δ . Fix any integer n ≥ n(i) (δ ) and consider the event ˜ ˜ (i) (i) (iii) (iv) ˜ (1) Hn (δ ) = Gmn ∩ Hmn (δ ) ∩ Hmn ∩ Hmn (γ /16) ∩ Hmn . ˜ ˜ ˜ ˜ ˜ 1562 , ACTIVIZED L EARNING (1) ˜ By Lemma 49 and the fact that mk ≥ mn for all k ≥ 1, since n ≥ n(i) (δ ) ≥ 24τ (1/6; δ ), on Hn (δ ), ˇ ˜ ˜f , ∀m ∈ Uk , ˆ ∀k ∈ 1, . . . , d ⋆ ⋆ ˆ (k) ˆ (k) ∆m Xm ,W2 ,Vm−1 ≤ (3/2)∆m Xm ,W2 ,Vmk . ˇ (67) Now fix any k ∈ 1, . . . , d˜f . Since n ≥ n(i) (δ ) ≥ 27·4k−1 , Lemma 50 implies Tk⋆ ≥ 18, which means ⋆ ≤ T ⋆ /4 . Let N = (4/3)∆(k) W ,W ,V ⋆ ˆ ˆ Uk , 3T ⋆ /4 − ⌈T ⋆ /4⌉ ≥ 4T ⋆ /9. Also note ˇ I 1 2 k k k m∈Uk mk k ˆ and note that Uk = Tk⋆ / ˆ (k) 3∆mk ˇ k ⋆ W1 ,W2 ,Vmk ˇ mk ˇ mk ˇ , so that Nk ≤ (4/9)Tk⋆ . Thus, we have ⋆ ˜ (1) P Hn (δ ) ∩ Imk > 3Tk⋆ /4 m=mk−1 +1 ˆ ⋆ ⋆ ˜ (1) ˜ (1) Imk > 4Tk⋆ /9 ≤ P Hn (δ ) ∩ Imk > Nk ≤ P Hn (δ ) ∩ ˆ ˆ m∈Uk m∈Uk ⋆ ˆm ˜ (1) ≤ P Hn (δ ) ∩ ½[2γ /3,∞) ∆(k) Xm ,W2 ,Vmk > Nk , ˇ mk ˆ (68) ˆ m∈Uk ⋆ ˜ ˇ where this last inequality is by (67). To simplify notation, define Zk = Tk⋆ , mk ,W1 ,W2 ,Vmk . By ˇ Lemmas 43 and 44 (with β = 3/32, ζ = 2γ /3, α = 3/4, and ξ = γ /16), since n ≥ n(i) (δ ) ≥ ˆ ˜ (1) 24 · max τ (iv) (γ /16; δ ), τ (3/32; δ ) , on Hn (δ ), ∀m ∈ Uk , ˜ ˇ ˇ p2γ /3 (k, mk , m) ≤ P (x : px (k, mk ) ≥ γ /2) + exp −γ 2 M(m)/256 ¯ ˜ ˇ ≤ P (x : px (k, mk ) ≥ γ /2) + exp −γ 2 M(mk )/256 ˇ (k) ⋆ ˆ ≤ ∆mk W1 ,W2 ,Vmk . ˇ ˇ (k) (1) ⋆ ˆ ˜n ˜n ˜ ˇ Letting G′ (k) denote the event p2γ /3 (k, mk , m) ≤ ∆mk W1 ,W2 ,Vmk , we see that G′ (k) ⊇ Hn (δ ). ¯ ˇ ˇ ⋆ ˆ (k) ˜ variables are conditionally independent given Zk for Thus, since the ½[2γ /3,∞) ∆m Xm ,W2 ,Vmk ˇ ˆ m ∈ Uk , each with respective conditional distribution Bernoulli p2γ /3 (k, mk , m) , the law of total ¯ ˇ probability and a Chernoff bound imply that (68) is at most ˜n P G′ (k) ∩ = E P ⋆ ˆm ½[2γ /3,∞) ∆(k) Xm ,W2 ,Vmk ˇ ˆ m∈Uk ⋆ ˆm ½[2γ /3,∞) ∆(k) Xm ,W2 ,Vmk ˇ ˆ m∈Uk (k) ⋆ ˆ ≤ E exp −∆mk W1 ,W2 ,Vmk ˇ ˇ ˆ Uk /27 > Nk ˜ > Nk Zk · ½G′n (k) ˜ ≤ E [exp{−Tk⋆ /162}] ≤ exp −n/ 243 · 4k−1 1563 , H ANNEKE ˜ ˜ ˜ (1) where the last inequality is by Lemma 50. Thus, there exists Gn (k) with P Hn (δ ) \ Gn (k) ≤ exp −n/ 243 · 4k−1 ˜ (i) ˜ (1) Hn (δ ) = Hn (δ ) ∩ (1) ˜ ˜ such that, on Hn (δ ) ∩ Gn (k), we have d˜f ˜ k=1 Gn (k), mk ˆ ⋆ m=mk−1 +1 Imk ˆ ≤ 3Tk⋆ /4. Defining a union bound implies ˜ ˜ (1) ˜ (i) P Hn (δ ) \ Hn (δ ) ≤ d˜f · exp −n/ 243 · 4d f −1 , (69) (i) ˆ ⋆ ⋆ ˜ and on Hn (δ ), every k ∈ 1, . . . , d˜f has mk mk−1 +1 Imk ≤ 3Tk⋆ /4. In particular, this means the Cmk m= ˆ ˆ ⋆ ˆ factors are redundant in Q⋆ , so that tk = mk mk−1 +1 Imk . k m= ˆ To get the stated probability bound, a union bound implies that (i) ˜ (1) 1 − P Hn (δ ) ≤ (1 − P (Hmn (δ ))) + 1 − P Hmn ˜ ˜ (iv) + 1 − P Hmn ˜ (i) (i) (iii) + P Hmn \ Hmn (γ /16) ˜ ˜ (i) + P Hmn \ Gmn ˜ ˜ ˜ ˜ ≤ δ + c(i) · exp −M (mn ) /4 ˜ ˜ ˜ + c(iii) (γ /16) · exp −M (mn ) γ 2 /256 + 3d˜f · exp {−2mn } −1 ˜ ˜ ˜ + 121d˜f δ f · exp −M (mn ) /60 ˜ ≤ δ + c(i) + c(iii) (γ /16) + 124d˜f δ f−1 · exp −mn δ f γ 2 /512 . ˜ ˜ (70) ˜ Since n ≥ n(i) (δ ) ≥ 24, we have mn ≥ n/48, so that summing (69) and (70) gives us ˜ ˜ ˜ ˜ (i) 1 − P Hn (δ ) ≤ δ + c(i) + c(iii) (γ /16) + 125d˜f δ f−1 · exp −nδ f γ 2 / 512 · 48 · 4d f −1 . (71) Finally, note that we have chosen n(i) (δ ) sufficiently large so that (71) is at most 2δ . The next lemma indicates that the redundancy of the “t < ⌊3Tk /4⌋” constraint, just established in Lemma 51, implies that all y labels obtained while k ≤ d˜f are consistent with the target function. ˆ Lemma 52 Consider running Meta-Algorithm 3 with a budget n ∈ N, while f is the target func˜ (ii) tion and P is the data distribution. There is an event Hn and (C, P, f , γ )-dependent constants (ii) (ii) (ii) (ii) ˜ (ii) ≤ δ , ˜ (i) c1 , c2 ∈ [1, ∞) such that, for any δ ∈ (0, 1), if n ≥ c1 ln c2 /δ , then P Hn (δ ) \ Hn ˜ ˜ ˜ ˜ ˜ ⋆ ˜ (i) ˜ (ii) and on Hn (δ ) ∩ Hn , we have V (d f ) = Vmd˜ = Vm ˜ . ˆ ˆ df f (ii) (i) Proof Define c1 = max c1 , r 192d , ˜ ˜ (1−γ )/6 211 ˜ 1/3 δf (ii) (i) , c2 = max c2 , r ˜ ˜ 8e (1−γ )/6 , c(ii) , exp {τ ∗ } , let n(ii) (δ ) = (ii) (ii) (ii) ˜ (ii) c1 ln c2 /δ , suppose n ≥ n(ii) (δ ), and define the event Hn = Hmn . ˜ ˜ ˜ ˜ (i) ˜ (ii) By Lemma 41, since n ≥ n(ii) (δ ) ≥ 24 · max {τ ((1 − γ )/6; δ ), τ ∗ }, on Hn (δ ) ∩ Hn , ∀m ∈ N ˜ and k ∈ 1, . . . , d˜f with either k = 1 or m > mn , ⋆ ⋆ ⋆ ˆ (k) ˆ (k) ˆ (k) ∆m Xm ,W2 ,Vm−1 < γ ⇒ Γm Xm , − f (Xm ),W2 ,Vm−1 < Γm Xm , f (Xm ),W2 ,Vm−1 . 1564 (72) ACTIVIZED L EARNING ˜ Recall that mn ≤ min {⌈T1 /4⌉ , 2n } = ⌈⌈2n/3⌉ /4⌉. Therefore, Vmn is obtained purely by mn exe˜ ˜ cutions of Step 8 while k = 1. Thus, for every m obtained in Meta-Algorithm 3, either k = 1 or m > mn . We now proceed by induction on m. We already know V0 = C = V0⋆ , so this serves as ˜ our base case. Now consider some value m ∈ N obtained in Meta-Algorithm 3 while k ≤ d˜f , and ⋆ suppose every m′ < m has Vm′ = Vm′ . But this means that Tk = Tk⋆ and the value of t upon obtaining m−1 ⋆ ⋆ ˆ (k) this particular m has t ≤ ℓ=mk−1 +1 Iℓk . In particular, if ∆m (Xm ,W2 ,Vm−1 ) ≥ γ , then Imk = 1, so ˆ ˆ ⋆ ⋆ ⋆ ˜ (ii) ˜ (i) that t < m mk−1 +1 Imk ; by Lemma 51, on Hn (δ ) ∩ Hn , m mk−1 +1 Imk ≤ mk mk−1 +1 Imk ≤ 3Tk⋆ /4, ℓ= ˆ ℓ= ˆ ℓ= ˆ ⋆ /4, and therefore y = Y = f (X ); this implies V = V ⋆ . On the other hand, on so that t < 3Tk ˆ m m m m (ii) (k) (i) ˆ ˜ ˜ Hn (δ ) ∩ Hn , if ∆m (Xm ,W2 ,Vm−1 ) < γ , then (72) implies ˆ (k) y = argmax Γm (Xm , y,W2 ,Vm−1 ) = f (Xm ), ˆ y∈{−1,+1} ⋆ ˜ (i) ˜ (ii) so that again Vm = Vm . Thus, by the principle of induction, on Hn (δ ) ∩ Hn , for every m ∈ N ˜f ) ⋆ ⋆ obtained while k ≤ d˜f , we have Vm = Vm ; in particular, this implies V (d = Vmd˜ = Vm ˜ . The bound ˆ ˆ df f ˜ (i) ˜ (ii) Hn (δ ) \ Hn on P then follows from Lemma 41, as we have chosen that (27) (with τ = mn ) is at most δ . ˜ n(ii) (δ ) sufficiently large so Lemma 53 Consider running Meta-Algorithm 3 with a budget n ∈ N, while f is the target func(iii) (iii) ˜ ˜ tion and P is the data distribution. There exist (C, P, f , γ )-dependent constants c1 , c2 ∈ [1, ∞) −3 ), λ ∈ [1, ∞), and n ∈ N, there exists an event H (iii) (δ , λ ) having ˜n such that, for any δ ∈ (0, e ˜ (ii) ˜ (iii) ˜ (i) P Hn (δ ) ∩ Hn \ Hn (δ , λ ) ≤ δ with the property that, if (iii) n ≥ c1 θ f (d/λ ) ln2 ˜ ˜ (i) (ii) (iii) c2 λ ˜ δ , (iii) ˜ ˜ ˜ then on Hn (δ ) ∩ Hn ∩ Hn (δ , λ ), at the conclusion of Meta-Algorithm 3, Ld˜f ≥ λ . (iii) Proof Let c1 ˜ (i) 10+2d˜f (ii) d·d˜f ·4 ˜ γ 3 δ f3 = max c1 , c1 , ˜ ˜ , r192d (3/32) (iii) , c2 ˜ (i) (ii) = max c2 , c2 , r 8e ˜ ˜ (3/32) , fix any δ ∈ (iii) (iii) ˜ ˜ ˜ (0, e−3 ), λ ∈ [1, ∞), let n(iii) (δ , λ ) = c1 θ f (d/λ ) ln2 (c2 λ /δ ), and suppose n ≥ n(iii) (δ , λ ). ˜ ˜ ˆ Define a sequence ℓi = 2i for integers i ≥ 0, and let ι = log2 42+d f λ /γ δ f . Also define ˜ ˆ φ (m, δ , λ ) = max {φ (m; δ /2ι ) , d/λ }, where φ is defined in Lemma 29. Then define the events ˜ H (3) (δ , λ ) = ˆ ι i=1 ˜ (iii) ˜ ˆ Hℓi (δ /2ι ) , Hn (δ , λ ) = H (3) (δ , λ ) ∩ md˜f ≥ ℓι . ˇ ˆ ˆ ˇ Note that ι ≤ n, so that ℓι ≤ 2n , and therefore the truncation in the definition of md˜f , which enforces ˆ ˜f · 2n + 1, mk−1 , will never be a factor in whether or not m ˜ ≥ ℓι is satisfied. ˇ df md˜f ≤ max d ˇ ˆ ˆ (ii) (ii) ⋆ ˜ (ii) ˆ ˜ (i) Since n ≥ n(iii) (λ , δ ) ≥ c1 ln c2 /δ , Lemma 52 implies that on Hn (δ ) ∩ Hn , Vmd˜ = Vm ˜ . ˜ ˜ ˆ f df Recall that this implies that all y values obtained while m ≤ md˜f are consistent with their respective ˆ ˆ 1565 H ANNEKE ⋆ ⋆ f (Xm ) values, so that every such m has Vm = Vm as well. In particular, Vmd˜ = Vm ˜ . Also note that ˇ ˇ df f n(iii) (δ , λ ) Thus, on 24 · τ (iv) (γ /16; δ ), ≥ so that ˜ (ii) ˜ (iii) ˜ (i) Hn (δ ) ∩ Hn ∩ Hn (δ , λ ), τ (iv) (γ /16; δ ) (taking ˆ ∆(k) ≤ mn , and recall we always have mn ≤ md˜f . ˜ ˜ ˇ as in Meta-Algorithm 3) ˜ ⋆ ˆ ˜ ˆ (d ) ∆(d f ) = ∆m ˜f W1 ,W2 ,Vmd˜ ˇ ˇ df (Lemma 52) f ˇ ˇ d˜ ≤ P x : px d˜f , md˜f ≥ γ /8 + 4m−1 (Lemma 44) f ˜ 8P d f ≤ ˜ γ P d f −1 ˜ ⋆ S d f Vm ˜ ˇ df ˜ S d f −1 ˜ ˜ ≤ 8/γ δ f P d f ⋆ Vm ˜ ˇd + 4md˜ ˇ −1 f ˜ ⋆ S d f Vmd˜ ˇ ˜ ˜ ˜ (Markov’s ineq.) f + 4md˜ ˇ −1 ˜ ˜ ≤ 8/γ δ f P d f S d f Vℓ⋆ ˆ ι (Lemma 35) f f ˜ (iii) (defn of Hn (δ , λ )) −1 + 4ℓι ˆ ˜ ˜ ˆ ≤ 8/γ δ f P d f S d f B f , φ (ℓι , δ , λ ) + 4ℓ−1 ˆ ι (Lemma 29) ˜ ˜ ˜ ˆ ≤ 8/γ δ f θ f (d/λ )φ (ℓι , δ , λ ) + 4ℓ−1 ˆ ι ˜ (defn of θ f (d/λ )) ˜ ˜ ˜ ˆ ≤ 12/γ δ f θ f (d/λ )φ (ℓι , δ , λ ) = ˜ ˆ (φ (ℓι , δ , λ ) ≥ ℓ−1 ) ˆ ι ˜ ˆ 12θ f (d/λ ) d ln (2e max {ℓι , d} /d) + ln (4ι /δ ) ˆ max 2 , d/λ . ˜ ℓι γδ f ˆ (73) ˆ Plugging in the definition of ι and ℓι , ˆ ˆ d ln (2e max {ℓι , d} /d) + ln (4ι /δ ) ˜ ˜ ˆ ˜ ˜ ≤ (d/λ )γ δ f 4−1−d f ln 41+d f λ /δ γ δ f ≤ (d/λ ) ln (λ /δ ) . ℓι ˆ ˜ ˜ Therefore, (73) is at most 24θ f (d/λ )(d/λ ) ln (λ /δ ) /γ δ f . Thus, since (i) (i) (ii) (ii) ˜ ˜ ˜ n(iii) (δ , λ ) ≥ max c1 ln c2 /δ , c1 ln c2 /δ ˜ , ˜ (i) ˜ (ii) ˜ (iii) Lemmas 51 and 52 imply that on Hn (δ ) ∩ Hn ∩ Hn (δ , λ ), ˜ ˆ Ld˜f = Td⋆f / 3∆(d f ) ˜ ˜ ˜ ˆ ≥ 41−d f 2n/ 9∆(d f ) ˜ ˜ 41−d f γ δ f n ≥ λ ln(λ /δ ) ≥ λ . ≥ ˜ 9 · 24 · θ f (d/λ )(d/λ ) ln (λ /δ ) ˜ (i) ˜ (ii) ˜ (iii) Now we turn to bounding P Hn (δ ) ∩ Hn \ Hn (δ , λ ) . By a union bound, we have ˆ ι ˜ 1 − P H (3) (δ , λ ) ≤ i=1 ˆ (1 − P (Hℓi (δ /2ι ))) ≤ δ /2. 1566 (74) ACTIVIZED L EARNING ˜ (i) ˜ (ii) ˜ Thus, it remains only to bound P Hn (δ ) ∩ Hn ∩ H (3) (δ , λ ) ∩ md˜f < ℓι ˇ ˆ . ˇ ˆ For each i ∈ {0, 1, . . . , ι − 1}, let Qi = ⋆ ˇ m ∈ (ℓi , ℓi+1 ] ∩ Ud˜f : Imd˜ = 1 . Now consider the set f ˇ ˆ I of all i ∈ {0, 1, . . . , ι − 1} with ℓi ≥ mn and (ℓi , ℓi+1 ] ∩ Ud˜f = ∅. Note that n(iii) (δ , λ ) ≥ 48, so that ˜ ℓ0 < mn . Fix any i ∈ I. Since n(iii) (λ , δ ) ≥ 24· τ (1/6; δ ), we have mn ≥ τ (1/6; δ ), so that Lemma 49 ˜ ˜ (i) (ii) (3) (δ , λ ), letting Q = 2 · 46+d˜f d/γ 2 δ 2 θ (d/λ ) ln(λ /δ ), ˜ ˜f ¯ ˜ ˜ ˜ implies that on Hn (δ ) ∩ Hn ∩ H f ˇ ¯ ˜ (ii) ˜ ˜ (i) P Hn (δ ) ∩ Hn ∩ H (3) (δ , λ ) ∩ Qi > Q W2 ,Vℓ⋆ i ˜ ˆ (d ) m ∈ (ℓi , ℓi+1 ] ∩ N : ∆m f Xm ,W2 ,Vℓ⋆ ≥ 2γ /3 i ≤P ¯ > Q W2 ,Vℓ⋆ . (75) i ˜ ˆm ½[2γ /3,∞) ∆(d f ) Xm ,W2 ,Vℓ⋆i are conditionally (given W2 ,Vℓ⋆ ) indepeni dent, each with respective conditional distribution Bernoulli with mean p2γ /3 d˜f , ℓi , m . Since ¯ n(iii) (δ , λ ) ≥ 24 · τ (3/32; δ ), we have mn ≥ τ (3/32; δ ), so that Lemma 43 (with ζ = 2γ /3, α = 3/4, ˜ (i) ˜ n (δ ) ∩ Hn ∩ H (3) (δ , λ ), each of these m values has ˜ (ii) ˜ and β = 3/32) implies that on H For m > ℓi , the variables ˜ p2γ /3 d˜f , ℓi , m ≤ P x : px d˜f , ℓi ≥ γ /2 + exp −M(m)γ 2 /256 ¯ ˜ ≤ ˜ 2P d f S d f Vℓ⋆ i ˜ γ P d f −1 ˜ S d f −1 ˜ Vℓ⋆ i ˜ ˜ ≤ 2/γ δ f P d f S d f Vℓ⋆ i ˜ + exp −M(ℓi )γ 2 /256 (Markov’s ineq.) ˜ + exp −M(ℓi )γ 2 /256 ˜ ˜ ˜ ˜ ≤ 2/γ δ f P d f S d f B f , φ (ℓi , δ , λ ) (Lemma 35) ˜ + exp −M(ℓi )γ 2 /256 ˜ ˜ ˜ ˜ ≤ 2/γ δ f θ f (d/λ )φ (ℓi , δ , λ ) + exp −M(ℓi )γ 2 /256 (Lemma 29) ˜ (defn of θ f (d/λ )). Denote the expression in this last line by pi , and let B(ℓi , pi ) be a Binomial(ℓi , pi ) random vari˜ (ii) ˜ ˜ (i) able. Noting that ℓi+1 − ℓi = ℓi , we have that on Hn (δ ) ∩ Hn ∩ H (3) (δ , λ ), (75) is at most ¯ P B(ℓi , pi ) > Q . Next, note that ˜ ˜ ˜ ˜ ℓi pi = (2/γ δ f )θ f (d/λ )ℓi φ (ℓi , δ , λ ) + ℓi · exp −ℓ3 δ f γ 2 /512 . i ˜ Since u · exp −u3 ≤ (3e)−1/3 for any u, letting u = ℓi δ f γ /8 we have ˜ ˜ ˜ ˜ ℓi · exp −ℓ3 δ f γ 2 /512 ≤ 8/γ δ f u · exp −u3 ≤ 8/ γ δ f (3e)1/3 ≤ 4/γ δ f . i 1567 H ANNEKE ˜ Therefore, since φ (ℓi , δ , λ ) ≥ ℓ−1 , we have that ℓi pi is at most i ˆ 6 ˜ 4ι 6 ˜ ˜ θ (d/λ )ℓi φ (ℓi , δ , λ ) ≤ θ (d/λ ) max 2d ln (2eℓι ) + 2 ln , ℓι d/λ ˆ ˆ ˜f f ˜f f δ γδ γδ 6 ˜ ≤ θ (d/λ ) max 2d ln ˜ f γδ f ≤ 6 ˜ θ (d/λ ) max 4d ln ˜ f γδ f ˜ ˜ 43+d f eλ ˜ γδ f + 2 ln ˜ d44+d f λ 6 ˜ ln θ f (d/λ ) · ≤ ˜f ˜f δ γδ γδ ˜ d43+d f , ˜ γδ f ˜ ˜ 43+d f λ ˜ γδ f δ 43+d f 2λ ˜ γδ f δ , d43+d f ˜ γδ f ˜ 46+d f d ˜ λ ≤ θ f (d/λ ) ln ˜ 2δ 2 δ γ f ¯ = Q/2. (i) ¯ ¯ ˜ ˆ Therefore, a Chernoff bound implies P B(ℓi , pi ) > Q ≤ exp −Q/6 ≤ δ /2ι , so that on Hn (δ ) ∩ (ii) (3) (δ , λ ), (75) is at most δ /2ι . The law of total probability implies there exists an event ˜ ˜ ˆ Hn ∩ H (4) (i) (ii) ˜ ˜ ˜ ˜ ˜ (4) ˜ (i) ˆ Hn (i, δ , λ ) with P Hn (δ ) ∩ Hn ∩ H (3) (δ , λ ) \ Hn (i, δ , λ ) ≤ δ /2ι such that, on Hn (δ ) ∩ ˇ ¯ ˜ (4) ˜ (ii) ˜ Hn ∩ H (3) (δ , λ ) ∩ Hn (i, δ , λ ), Qi ≤ Q. Note that ˜ ˜ ˜ ˜ ˜ ˆ¯ ι Q ≤ log2 42+d f λ /γ δ f · 47+d f d/γ 2 δ f2 θ f (d/λ ) ln(λ /δ ) ˜ ˜ ˜ ˜ ≤ d˜f 49+d f /γ 3 δ f3 d θ f (d/λ ) ln2 (λ /δ ) ≤ 41−d f n/12. ⋆ m≤2mn Imd˜f ˜ ˜ (4) i∈I Hn (i, δ , λ ), Since (76) ˜ (i) ˜ (ii) ˜ ≤ n/12, if d˜f = 1 then (76) implies that on the event Hn (δ )∩ Hn ∩ H (3) (δ , λ )∩ ˇ ˆ¯ ≤ n/12 + i∈I Qi ≤ n/12 + ι Q ≤ n/6 ≤ ⌈T1⋆ /4⌉, so that m1 ≥ ℓι . ˇ ˆ (i) ˇ ˇ ˇ ˜ ˜ Otherwise, if d˜f > 1, then every m ∈ Ud˜f has m > 2mn , so that i≤ˆ Qi = i∈I Qi ; thus, on Hn (δ )∩ ι ˜f ˜f (4) (ii) ˇ ˜ ˜ ˜ ˆ¯ Hn ∩ H (3) (δ , λ ) ∩ i∈I Hn (i, δ , λ ), i∈I Qi ≤ ι Q ≤ 41−d n/12; Lemma 50 implies 41−d n/12 ≤ ⋆ m≤ℓι Im1 ˆ ˇ Td⋆ /4 , so that again we have md˜f ≥ ℓι . Combined with a union bound, this implies ˆ ˜ f ˜ (ii) ˜ ˜ (i) ˇ P Hn (δ ) ∩ Hn ∩ H (3) (δ , λ ) ∩ md˜f < ℓι ˆ ˜ (ii) ˜ ˜ (i) ≤ P Hn (δ ) ∩ Hn ∩ H (3) (δ , λ ) \ ≤ P i∈I ˜ (4) Hn (i, δ , λ ) i∈I ˜ (i) ˜ (ii) ˜ ˜ (4) Hn (δ ) ∩ Hn ∩ H (3) (δ , λ ) \ Hn (i, δ , λ ) ≤ δ /2. (77) ˜ (ii) ˜ (iii) ˜ (i) Therefore, P Hn (δ ) ∩ Hn \ Hn (δ , λ ) ≤ δ , obtained by summing (77) and (74). Proof [Theorem 16] If Λ p (ε /4, f , P) = ∞ then the result trivially holds. Otherwise, suppose (i) (ii) (iii) ε ∈ (0, 10e−3 ), let δ = ε /10, λ = Λ p (ε /4, f , P), c2 = max 10c2 , 10c2 , 10c2 , 10e(d + 1) , ˜ ˜ ˜ ˜ (i) (ii) (iii) and c1 = max c1 , c1 , c1 , 2 · 63 (d + 1)d˜ln(e(d + 1)) , and consider running Meta-Algorithm ˜ ˜ ˜ ˜ 1568 ACTIVIZED L EARNING 3 with passive algorithm A p and budget n ≥ c1 θ f (d/λ ) ln2 (c2 λ /ε ), while f is the target func˜ ˜ ˜ (i) ˜ (ii) ˜ (iii) ˜ n (δ ) ∩ Hn ∩ Hn (δ , λ ), Lemma 53 imtion and P is the data distribution. On the event H ˜ ⋆ plies Ld˜f ≥ λ , while Lemma 52 implies V (d f ) = Vm ˜ ; recalling that Lemma 35 implies that ˆ df ⋆ Vm ˜ ˆd f ˆ = ∅ on this event, we must have erLd˜ ( f ) = 0. Furthermore, if h is the classifier returned f ˆ by Meta-Algorithm 3, then Lemma 34 implies that er(h) is at most 2 er(A p (Ld˜f )), on a high ˆ ˜ (i) ˜ (ii) ˜ (iii) ˆ ˆ probability event (call it E2 in this context). Letting E3 (δ ) = E2 ∩ Hn (δ ) ∩ Hn ∩ Hn (δ , λ ), ˆ a union bound implies the total failure probability 1 − P(E3 (δ )) from all of these events is at most 4δ + e(d + 1) · exp −⌊n/3⌋/ 72d˜f (d + 1) ln(e(d + 1)) ≤ 5δ = ε /2. Since, for ℓ ∈ N with P Ld˜f = ℓ > 0, the sequence of Xm values appearing in Ld˜f are conditionally distributed as P ℓ given |Ld˜f | = ℓ, and this is the same as the (unconditional) distribution of {X1 , X2 , . . . , Xℓ }, we have that ˆ E er h ≤ E 2 er A p Ld˜f ½E3 (δ ) + ε /2 = E E 2 er A p Ld˜f ˆ ≤2 sup ℓ≥Λ p (ε /4, f ,P) ½E3 (δ ) |Ld˜f | + ε /2 ˆ E [er(A p (Zℓ ))] + ε /2 ≤ ε . To specialize to the specific variant of Meta-Algorithm 3 stated in Section 5.2, take γ = 1/2. Appendix E. Proofs Related to Section 6: Agnostic Learning This appendix contains the proofs of our results on learning with noise. Specifically, Appendix E.1 provides the proof of the counterexample from Theorem 22, demonstrating that there is no activizer ˇ for the A p passive learning algorithm described in Section 6.2 in the agnostic case. Appendix E.2 presents the proof of Lemma 26 from Section 6.7, bounding the label complexity of Algorithm 5 under Condition 1. Finally, Appendix E.3 presents a proof of Theorem 28, demonstrating that any active learning algorithm can be modified to trivialize the misspecified model case. The notation used throughout Appendix E is taken from Section 6. E.1 Proof of Theorem 22: Negative Result for Agnostic Activized Learning ˇ It suffices to show that A p achieves a label complexity Λ p such that, for any label complexity Λa achieved by any active learning algorithm Aa , there exists a distribution PXY on X × {−1, +1} such that PXY ∈ Nontrivial(Λ p ; C) and yet Λa (ν + cε , PXY ) = o (Λ p (ν + ε , PXY )) for every constant c ∈ (0, ∞). Specifically, we will show that there is a distribution PXY for which Λ p (ν + ε , PXY ) = Θ(1/ε ) and Λa (ν + ε , PXY ) = o(1/ε ). Let P({0}) = 1/2, and for any measurable A ⊆ (0, 1], P(A) = λ (A)/2, where λ is Lebesgue measure. Let D be the family of distributions PXY on X × {−1, +1} characterized by the properties that the marginal distribution on X is P, η (0; PXY ) ∈ (1/8, 3/8), and ∀x ∈ (0, 1], η (x; PXY ) = η (0; PXY ) + (x/2) · (1 − η (0; PXY )) . η (0;PXY Thus, η (x; PXY ) is a linear function. For any PXY ∈ D, since the point z ∗ = 1−2η (0;PXY )) has 1− η (z ∗ ; PXY ) = 1/2, we see that f = hz ∗ is a Bayes optimal classifier. Furthermore, for any η0 ∈ 1569 H ANNEKE [1/8, 3/8], |η (0; PXY ) − η0 | 1 − 2η0 1 − 2η (0; PXY ) = , − 1 − η0 1 − η (0; PXY ) (1 − η0 )(1 − η (0; PXY )) and since (1 − η0 )(1 − η (0; PXY )) ∈ (25/64, 49/64) ⊂ (1/3, 1), the value z = 1−2η0 1−η0 satisfies |η0 − η (0; PXY )| ≤ |z − z ∗ | ≤ 3|η0 − η (0; PXY )|. (78) Also note that under PXY , since (1 − 2η (0; PXY )) = (1 − η (0; PXY ))z ∗ , any z ∈ (0, 1) has er(hz ) − er(hz ∗ ) = z∗ z 1 − 2η (x; PXY ) dx = = (1 − η (0; PXY )) z∗ z z∗ z 1 − 2η (0; PXY ) − x(1 − η (0; PXY )) dx (z ∗ − x) dx = (1 − η (0; PXY )) ∗ (z − z)2 , 2 so that 5 7 (z − z ∗ )2 ≤ er(hz ) − er(hz ∗ ) ≤ (z − z ∗ )2 . 16 16 Finally, note that any x, x′ ∈ (0, 1] with |x − z ∗ | < |x′ − z ∗ | has (79) |1 − 2η (x; PXY )| = |x − z ∗ |(1 − η (0; PXY )) < |x′ − z ∗ |(1 − η (0; PXY )) = |1 − 2η (x′ ; PXY )|. ′ ′ ′ Thus, for any q ∈ (0, 1/2], there exists zq ∈ [0, 1] such that z ∗ ∈ [zq , zq + 2q] ⊆ [0, 1], and the clas′ ′ sifier h′ (x) = hz ∗ (x) · 1 − 2½(zq ,zq +2q] (x) has er(h) ≥ er(h′ ) for every classifier h with h(0) = q q ′ −1 and P(x : h(x) = hz ∗ (x)) = q. Noting that er(h′ ) − er(hz ∗ ) = limz↓zq er(hz ) − er(hz ∗ ) + q ′ er(hzq +2q ) − er(hz ∗ ) , (79) implies that er(h′ ) − er(hz ∗ ) ≥ q 5 16 ′ zq − z ∗ 2 ′ + zq + 2q − z ∗ 2 , and 5 ′ ′ since max{z ∗ − zq , zq + 2q − z ∗ } ≥ q, this is at least 16 q2 . In general, any h with h(0) = +1 has er(h) − er(hz ∗ ) ≥ 1/2 − η (0; PXY ) > 1/8 ≥ (1/8)P(x : h(x) = hz ∗ (x))2 . Combining these facts, we see that any classifier h has er(h) − er(hz ∗ ) ≥ (1/8)P (x : h(x) = hz ∗ (x))2 . (80) ˇ Lemma 54 The passive learning algorithm A p achieves a label complexity Λ p such that, for every PXY ∈ D, Λ p (ν + ε , PXY ) = Θ(1/ε ). ˇ ˆ Proof Consider the values η0 and z from A p (Zn ) for some n ∈ N. Combining (78) and (79), ˆ 7 ∗ )2 ≤ 63 (η − η (0; P ))2 ≤ 4(η − η (0; P ))2 . Let N = ˆ0 ˆ0 we have er(hz ) − er(hz ∗ ) ≤ 16 (ˆ − z z XY XY n ˆ 16 −1 ¯ 0 = Nn |{i ∈ {1, . . . , n} : Xi = 0,Yi = +1}| if Nn > 0, or η0 = 0 if ¯ |{i ∈ {1, . . . , n} : Xi = 0}|, and η ˆ ˆ ¯ Nn = 0. Note that η0 = η0 ∨ 1 ∧ 3 , and since η (0; PXY ) ∈ (1/8, 3/8), we have |η0 − η (0; PXY )| ≤ 8 8 ¯ |η0 − η (0; PXY )|. Therefore, for any PXY ∈ D, ¯ ˆ E [er(hz ) − er(hz ∗ )] ≤ 4E (η0 − η (0; PXY ))2 ≤ 4E (η0 − η (0; PXY ))2 ˆ ¯ ≤ 4E E (η0 − η (0; PXY ))2 Nn ½[n/4,n] (Nn ) + 4P(Nn < n/4). (81) ¯ By a Chernoff bound, P(Nn < n/4) ≤ exp{−n/16}, and since the conditional distribution of Nn η0 given Nn is Binomial(Nn , η (0; PXY )), (81) is at most 4E 16 68 4 15 1 < . η (0; PXY )(1 − η (0; PXY )) + 4 · exp {−n/16} ≤ 4 · · + 4 · Nn ∨ n/4 n 64 n n 1570 ACTIVIZED L EARNING ˇ For any n ≥ ⌈68/ε ⌉, this is at most ε . Therefore, A p achieves a label complexity Λ p such that, for any PXY ∈ D, Λ p (ν + ε , PXY ) = ⌈68/ε ⌉ = Θ(1/ε ). Next we establish a corresponding lower bound for any active learning algorithm. Note that this requires more than a simple minimax lower bound, since we must have an asymptotic lower bound for a fixed PXY , rather than selecting a different PXY for each ε value; this is akin to the strong minimax lower bounds proven by Antos and Lugosi (1998) for passive learning in the realizable case. For this, we proceed by reduction from the task of estimating a binomial mean; toward this end, the following lemma will be useful. Lemma 55 For any nonempty (a, b) ⊂ [0, 1], and any sequence of estimators pn : {0, 1}n → [0, 1], ˆ there exists p ∈ (a, b) such that, if B1 , B2 , . . . are independent Bernoulli(p) random variables, also independent from every pn , then E ( pn (B1 , . . . , Bn ) − p)2 = o(1/n). ˆ ˆ Proof We first establish the claim when a = 0 and b = 1. For any p ∈ [0, 1], let B1 (p), B2 (p), . . . be i.i.d. Bernoulli(p) random variables, independent from any internal randomness of the pn estiˆ mators. We proceed by reduction from hypothesis testing, for which there are known lower bounds. Specifically, it is known (e.g., Wald, 1945; Bar-Yossef, 2003) that for any p, q ∈ (0, 1), δ ∈ (0, e−1 ), any (possibly randomized) q : {0, 1}n → {p, q}, and any n ∈ N, ˆ n< (1 − 8δ ) ln(1/8δ ) =⇒ 8KL(p q) max P (q(B1 (p∗ ), . . . , Bn (p∗ )) = p∗ ) > δ , ˆ p∗ ∈{p,q} where KL(p q) = p ln(p/q) + (1 − p) ln((1 − p)/(1 − q)). It is also known (e.g., Poland and Hutter, 2006) that for p, q ∈ [1/4, 3/4], KL(p q) ≤ (8/3)(p − q)2 . Combining this with the above fact, we have that for p, q ∈ [1/4, 3/4], max P (q(B1 (p∗ ), . . . , Bn (p∗ )) = p∗ ) ≥ (1/16) · exp −128(p − q)2 n/3 . ˆ p∗ ∈{p,q} (82) Given the estimator pn from the lemma statement, we construct a sequence of hypothesis tests as ˆ follows. For i ∈ N, let αi = exp −2i and ni = 1/αi2 . Define p∗ = 1/4, and for i ∈ N, induc0 tively define qi (b1 , . . . , bni ) = argmin p∈{p∗ ,p∗ +αi } | pni (b1 , . . . , bni ) − p| for b1 , . . . , bni ∈ {0, 1}, and ˆ ˆ i−1 i−1 ˆ p∗ = argmax p∈{p∗ ,p∗ +αi } P (qi (B1 (p), . . . , Bni (p)) = p). Finally, define p∗ = limi→∞ p∗ . Note that i i i−1 i−1 ∞ ∗ < 1/2, p∗ , p∗ + α ∈ [1/4, 3/4], and 0 ≤ p∗ − p∗ ≤ 2 ∀i ∈ N, pi i i i−1 i−1 j=i+1 α j < 2αi+1 = 2αi . We generally have 1 E ( pni (B1 (p∗ ), . . . , Bni (p∗ )) − p∗ )2 ≥ E ( pni (B1 (p∗ ), . . . , Bni (p∗ )) − p∗ )2 − (p∗ − p∗ )2 ˆ ˆ i i 3 1 ˆ ≥ E ( pni (B1 (p∗ ), . . . , Bni (p∗ )) − p∗ )2 − 4αi4 . i 3 Furthermore, note that for any m ∈ {0, . . . , ni }, (p∗ )m (1 − p∗ )ni −m ≥ (p∗ )m (1 − p∗ )ni −m i i 1 − p∗ 1 − p∗ i ≥ 1 − 4αi2 1571 ni ≥ ni 1 − p∗ − 2αi2 i 1 − p∗ i ni ≥ exp −8αi2 ni ≥ e−8 , H ANNEKE so that the probability mass function of (B1 (p∗ ), . . . , Bni (p∗ )) is never smaller than e−8 times that of (B1 (p∗ ), . . . , Bni (p∗ )), which implies (by the law of the unconscious statistician) i i ˆ E ( pni (B1 (p∗ ), . . . , Bni (p∗ )) − p∗ )2 ≥ e−8 E ( pni (B1 (p∗ ), . . . , Bni (p∗ )) − p∗ )2 . ˆ i i i i By a triangle inequality, we have E ( pni (B1 (p∗ ), . . . , Bni (p∗ )) − p∗ )2 ≥ ˆ i i i αi2 P (qi (B1 (p∗ ), . . . , Bni (p∗ )) = p∗ ) . ˆ i i i 4 By (82), this is at least αi2 (1/16) · exp −128αi2 ni /3 ≥ 2−6 e−43 αi2 . 4 Combining the above, we have E ( pni (B1 (p∗ ), . . . , Bni (p∗ )) − p∗ )2 ≥ 3−1 2−6 e−51 αi2 − 4αi4 ≥ 2−9 e−51 n−1 − 4n−2 . ˆ i i For i ≥ 5, this is larger than 2−11 e−51 n−1 . Since ni diverges as i → ∞, we have that i E ( pni (B1 (p∗ ), . . . , Bni (p∗ )) − p∗ )2 = o(1/n), ˆ which establishes the result for a = 0 and b = 1. To extend this result to general nonempty ranges (a, b), we proceed by reduction from the above problem. Specifically, suppose p′ ∈ (0, 1), and consider the following independent random variables (also independent from the Bi (p′ ) variables and pn estimators). For each i ∈ N, Ci1 ∼ ˆ Bernoulli(a), Ci2 ∼ Bernoulli((b − a)/(1 − a)). Then for bi ∈ {0, 1}, define B′ (bi ) = max{Ci1 ,Ci2 · i bi }. For any given p′ ∈ (0, 1), the random variables B′ (Bi (p′ )) are i.i.d. Bernoulli (p), with p = i a + (b − a)p′ ∈ (a, b) (which forms a bijection between (0, 1) and (a, b)). Defining p′ (b1 , . . . , bn ) = ˆn ′ (b ), . . . , B′ (b )) − a)/(b − a), we have ( pn (B1 1 ˆ n n E ( pn (B1 (p), . . . , Bn (p)) − p)2 = (b − a)2 · E ˆ p′ (B1 (p′ ), . . . , Bn (p′ )) − p′ ˆn 2 . (83) We have already shown there exists a value of p′ ∈ (0, 1) such that the right side of (83) is not o(1/n). Therefore, the corresponding value of p = a + (b − a)p′ ∈ (a, b) has the left side of (83) not o(1/n), which establishes the result. We are now ready for the lower bound result for our setting. Lemma 56 For any label complexity Λa achieved by any active learning algorithm Aa , there exists a PXY ∈ D such that Λa (ν + ε , PXY ) = o(1/ε ). Proof The idea here is to reduce from the task of estimating the mean of iid Bernoulli trials, corresponding to the Yi values. Specifically, consider any active learning algorithm Aa ; we use Aa to construct an estimator for the mean of iid Bernoulli trials as follows. Suppose we have B1 , B2 , . . . , Bn i.i.d. Bernoulli(p), for some p ∈ (1/8, 3/8) and n ∈ N. We take the sequence of X1 , X2 , . . . random 1572 ACTIVIZED L EARNING variables i.i.d. with distribution P defined above (independent from the B j variables). For each i, we additionally have a random variable Ci with conditional distribution Bernoulli(Xi /2) given Xi , where the Ci are conditionally independent given the Xi sequence, and independent from the Bi sequence as well. We run Aa with this sequence of Xi values. For the t th label request made by the algorithm, say for the Yi value corresponding to some Xi , if it has previously requested this Yi already, then we simply repeat the same answer for Yi again, and otherwise we return to the algorithm the value 2 max{Bt ,Ci } − 1 for Yi . Note that in the latter case, the conditional distribution of max{Bt ,Ci } is Bernoulli(p + (1 − p)Xi /2), given the Xi that Aa requests the label of; thus, the Yi response has the same conditional distribution given Xi as it would have for the PXY ∈ D with η (0; PXY ) = p (i.e., η (Xi ; PXY ) = p + (1 − p)Xi /2). Since this Yi value is conditionally (given Xi ) independent from the previously returned labels and X j sequence, this is distributionally equivalent to running Aa under the PXY ∈ D with η (0; PXY ) = p. ˆ Let hn be the classifier returned by Aa (n) in the above context, and let zn denote the value ˆ 1−zn ˆ ˆ of z ∈ [2/5, 6/7] with minimum P(x : hz (x) = hn (x)). Then define pn = 2−zn ∈ [1/8, 3/8] and ˆ ˆ z z ∗ = 1−2p ∈ (2/5, 6/7). By a triangle inequality, we have |ˆ n − z ∗ | = 2P(x : hzn (x) = hz ∗ (x)) ≤ ˆ 1−p ˆ n (x) = hz ∗ (x)). Combining this with (80) and (78) implies that 4P(x : h 1 ˆ ˆ er(hn ) − er(hz ∗ ) ≥ P x : hn (x) = hz ∗ (x) 8 2 ≥ 1 1 (ˆ n − z ∗ )2 ≥ z ( pn − p)2 . ˆ 128 128 (84) In particular, by Lemma 55, we can choose p ∈ (1/8, 3/8) so that E ( pn − p)2 = o(1/n), which, by ˆ ˆ (84), implies E er(hn ) − ν = o(1/n). This means there is an increasing infinite sequence of values ˆ nk ∈ N, and a constant c ∈ (0, ∞) such that ∀k ∈ N, E er(hnk ) − ν ≥ c/nk . Supposing Aa achieves label complexity Λa , and taking the values εk = c/(2nk ), we have Λa (ν + εk , PXY ) > nk = c/(2εk ). Since εk > 0 and approaches 0 as k → ∞, we have Λa (ν + ε , PXY ) = o(1/ε ). Proof [of Theorem 22] The result follows from Lemmas 54 and 56. E.2 Proof of Lemma 26: Label Complexity of Algorithm 5 The proof of Lemma 26 essentially runs parallel to that of Theorem 16, with variants of each lemma from that proof adapted to the noise-robust Algorithm 5. As before, in this section we will fix a particular joint distribution PXY on X × {−1, +1} with marginal P on X , and then analyze the label complexity achieved by Algorithm 5 for that particular distribution. For our purposes, we will suppose PXY satisfies Condition 1 for some finite parameters µ and κ . We also fix any f ∈ cl(C(ε )). Furthermore, we will continue using the notation of ε >0 ⋆ Appendix B, such as S k (H), etc., and in particular we continue to denote Vm = {h ∈ C : ∀ℓ ≤ ⋆ m, h(Xℓ ) = f (Xℓ )} (though note that in this case, we may sometimes have f (Xℓ ) = Yℓ , so that Vm = C[Zm ]). As in the above proofs, we will prove a slightly more general result in which the “1/2” threshold in Step 5 can be replaced by an arbitrary constant γ ∈ (0, 1). ˆ For the estimators P4m used in the algorithm, we take the same definitions as in Appendix B.1. To be clear, we assume the sequences W1 and W2 mentioned there are independent from the entire 1573 H ANNEKE (X1 ,Y1 ), (X2 ,Y2 ), . . . sequence of data points; this is consistent with the earlier discussion of how these W1 and W2 sequences can be constructed in a preprocessing step. We will consider running Algorithm 5 with label budget n ∈ N and confidence parameter δ ∈ ˆ ˆ ˆ (0, e−3 ), and analyze properties of the internal sets Vi . We will denote by Vi , Li , and ik , the final values of Vi , Li , and ik , respectively, for each i and k in Algorithm 5. We also denote by m(k) ˆ ˆ and V (k) the final values of m and Vik +1 , respectively, obtained while k has the specified value in ˆ ˆ ˆ Algorithm 5; V (k) may be smaller than Viˆk when m(k) is not a power of 2. Additionally, define ⋆ = {(X ,Y )}2i Li m m m=2i−1 +1 . After establishing a few results concerning these, we will show that for n satisfying the condition in Lemma 26, the conclusion of the lemma holds. First, we have a few auxiliary definitions. For H ⊆ C, and any i ∈ N, define φi (H) = E sup h1 ,h2 ∈H er(h1 ) − erL⋆ (h1 ) − er(h2 ) − erL⋆ (h2 ) i i ˜ ˜ and Ui (H, δ ) = min K φi (H) + diam(H) ln(32i2 /δ ) ln(32i2 /δ ) + 2i−1 2i−1 ,1 , ˜ where for our purposes we can take K = 8272. It is known (see, e.g., Massart and N´ d´ lec, 2006; e e Gin´ and Koltchinskii, 2006) that for some universal constant c′ ∈ [2, ∞), e φi+1 (H) ≤ c′ max diam(H)2−i d log2 2 , 2−i di . diam(H) (85) We also generally have φi (H) ≤ 2 for every i ∈ N. The next lemma is taken from the work of Koltchinskii (2006) on data-dependent Rademacher complexity bounds on the excess risk. Lemma 57 For any δ ∈ (0, e−3 ), any H ⊆ C with f ∈ cl(H), and any i ∈ N, on an event Ki with P(Ki ) ≥ 1 − δ /4i2 , ∀h ∈ H, ˆ erL⋆ (h) − min erL⋆ (h′ ) ≤ er(h) − er( f ) + Ui (H, δ ) i i ′ h ∈H ˆ er(h) − er( f ) ≤ erL⋆ (h) − erL⋆ ( f ) + Ui (H, δ ) i i ˜ ˆ min Ui (H, δ ), 1 ≤ Ui (H, δ ). Lemma 57 essentially follows from a version of Talagrand’s inequality. The details of the proof may be extracted from the proofs of Koltchinskii (2006), and related derivations have previously been presented by Hanneke (2011) and Koltchinskii (2010). The only minor twist here is that f need only be in cl(H), rather than in H itself, which easily follows from Koltchinskii’s original results, since the Borel-Cantelli lemma implies that with probability one, every ε > 0 has some g ∈ H(ε ) (very close to f ) with erL⋆ (g) = erL⋆ ( f ). i i For our purposes, the important implications of Lemma 57 are summarized by the following lemma. Lemma 58 For any δ ∈ (0, e−3 ) and any n ∈ N, when running Algorithm 5 with label budget n and ˆ ˆ confidence parameter δ , on an event Jn (δ ) with P(Jn (δ )) ≥ 1 − δ /2, ∀i ∈ {0, 1, . . . , id+1 }, if V2⋆i ⊆ Vi 1574 ACTIVIZED L EARNING ˆ then ∀h ∈ Vi , ˆ ˆ erL⋆ (h) − min erL⋆ (h′ ) ≤ er(h) − er( f ) + Ui+1 (Vi , δ ) i+1 i+1 ˆ h′ ∈Vi ˆ ˆ er(h) − er( f ) ≤ erL⋆ (h) − erL⋆ ( f ) + Ui+1 (Vi , δ ) i+1 i+1 ˆ ˆ ˜ ˆ min Ui+1 (Vi , δ ), 1 ≤ Ui+1 (Vi , δ ). ˆ Proof For each i, consider applying Lemma 57 under the conditional distribution given Vi . The ⋆ is independent from V , as are the Rademacher variables in the definition of R ˆi ˆ i+1 (Vi ). Furˆ set Li+1 thermore, by Lemma 35, on H ′ , f ∈ cl V2⋆i , so that the conditions of Lemma 57 hold. The law of total probability then implies the existence of an event Ji of probability P(Ji ) ≥ 1 − δ /4(i + 1)2 , on ˆ which the claimed inequalities hold for that value of i if i ≤ id+1 . A union bound over values of i then implies the existence of an event Jn (δ ) = i Ji with probability P(Jn (δ )) ≥ 1 − i δ /4(i + 1)2 ≥ ˆ 1 − δ /2 on which the claimed inequalities hold for all i ≤ id+1 . Lemma 59 For some (C, PXY , γ )-dependent constants c, c∗ ∈ [1, ∞), for any δ ∈ (0, e−3 ) and integer n ≥ c∗ ln(1/δ ), when running Algorithm 5 with label budget n and confidence parameter δ , on (i) (ii) ˆ event Jn (δ ) ∩ Hn ∩ Hn , every i ∈ {0, 1, . . . , id˜f } satisfies V2⋆i di + ln(1/δ ) ˆ ⊆ Vi ⊆ C c 2i κ 2κ −1 , ˜ ˆ and furthermore V ⋆(d˜f ) ⊆ V (d f ) . m ˆ ˜ √ Proof Define c = 24Kc′ µ 2κ 2κ −1 , c∗ = max τ ∗ , 8d µ c1/κ r(1−γ )/6 1 2κ −1 log2 4µ c1/κ r(1−γ )/6 , and suppose n ≥ c∗ ln(1/δ ). We now proceed by induction. As the right side equals C for i = 0, the claimed ˆ inclusions are certainly true for V0 = C, which serves as our base case. Now suppose some i ∈ ˆd˜ } satisfies {0, 1, . . . , i f V2⋆i di + ln(1/δ ) ˆ ⊆ Vi ⊆ C c 2i κ 2κ −1 . (86) In particular, Condition 1 implies di + ln(1/δ ) ˆ diam(Vi ) ≤ diam C c 2i κ 2κ −1 ≤ µc 1 κ di + ln(1/δ ) 2i 1 2κ −1 . (87) ˆ ˆ ˆ If i < id˜f , then let k be the integer for which ik−1 ≤ i < ik , and otherwise let k = d˜f . Note that we ˆ certainly have i1 ≥ ⌊log2 (n/2)⌋, since m = ⌊n/2⌋ ≥ 2⌊log2 (n/2)⌋ is obtained while k = 1. Therefore, if k > 1, di + ln(1/δ ) 4d log2 (n) + 4 ln(1/δ ) ≤ , 2i n 1575 H ANNEKE so that (87) implies 1 2κ −1 4d log2 (n) + 4 ln(1/δ ) n 1 ˆ diam Vi ≤ µ c κ . By our choice of c∗ , the right side is at most r(1−γ )/6 . Therefore, since Lemma 35 implies f ∈ cl V2⋆i (i) ˆ ˆ on Hn , we have Vi ⊆ B f , r(1−γ )/6 when k > 1. Combined with (86), we have that V2⋆i ⊆ Vi , and ˆ either k = 1, or Vi ⊆ B( f , r(1−γ )/6 ) and 4m > 4⌊n/2⌋ ≥ n. Now consider any m with 2i + 1 ≤ m ≤ ˜ ⋆ min 2i+1 , m(d f ) , and for the purpose of induction suppose Vm−1 ⊆ Vi+1 upon reaching Step 5 for ˆ ˆ that value of m in Algorithm 5. Since Vi+1 ⊆ Vi and n ≥ τ ∗ , Lemma 41 (with ℓ = m − 1) implies that (i) (ii) on Hn ∩ Hn , ˆ (k) ˆ (k) ˆ (k) ∆4m (Xm ,W2 ,Vi+1 ) < γ =⇒ Γ4m (Xm , − f (Xm ),W2 ,Vi+1 ) < Γ4m (Xm , f (Xm ),W2 ,Vi+1 ) , ⋆ ⋆ so that after Step 8 we have Vm ⊆ Vi+1 . Since (86) implies that the Vm−1 ⊆ Vi+1 condition holds if i + 1 (at which time V ˆ Algorithm 5 reaches Step 5 with m = 2 i+1 = Vi ), we have by induction that (i) (ii) i+1 , m(d˜f ) . This establishes the ⋆ ⊆V ˆ on Hn ∩ Hn , Vm i+1 upon reaching Step 9 with m = min 2 final claim of the lemma, given that the first claim holds. For the remainder of this inductive proof, ˆ suppose i < id˜f . Since Step 8 enforces that, upon reaching Step 9 with m = 2i+1 , every h1 , h2 ∈ Vi+1 (i) (ii) have erLi+1 (h1 ) − erLi+1 (h2 ) = erL⋆ (h1 ) − erL⋆ (h2 ), on Jn (δ ) ∩ Hn ∩ Hn we have ˆ ˆ i+1 i+1 ˆ Vi+1 ⊆ ˆ ˆ ˆ h ∈ Vi : erL⋆ (h) − ′min erL⋆ (h′ ) ≤ Ui+1 Vi , δ i+1 i+1 ⋆ h ∈V 2i+1 ˆ ˆ ˆ ⊆ h ∈ Vi : erL⋆ (h) − erL⋆ ( f ) ≤ Ui+1 Vi , δ i+1 i+1 ˆ ˆ ˆ ⊆ Vi ∩ C 2Ui+1 Vi , δ ˆ ˜ ⊆ C 2Ui+1 Vi , δ , (88) where the second line follows from Lemma 35 and the last two inclusions follow from Lemma 58. ˆ Focusing on (88), combining (87) with (85) (and the fact that φi+1 (Vi ) ≤ 2), we can bound the value ˆ ˜ i+1 Vi , δ as follows. of U 2 1 ˆ ln(32(i + 1) /δ ) ≤ √µ c 2κ diam(Vi ) i 2 ≤ √ µc di + ln(1/δ ) 2i 2di + 2 ln(1/δ ) 2i+1 1 2κ √ 1 ≤ 4 µ c 2κ ′√ ˆ φi+1 (Vi ) ≤ c µc ′√ ≤ 4c 1 2κ µc 1 2κ 1 4κ −2 ln(32(i + 1)2 /δ ) 2i 1 4κ −2 d(i + 1) + ln(1/δ ) 2i+1 di + ln(1/δ ) 2i 1 4κ −2 d(i + 1) + ln(1/δ ) 2i+1 1576 1 2 8(i + 1) + 2 ln(1/δ ) 2i+1 κ 2κ −1 , d(i + 2) 2i κ 2κ −1 , 1 2 1 2 ACTIVIZED L EARNING and thus d(i + 1) + ln(1/δ ) 2i+1 ˜ ˆ ˜ √ Ui+1 (Vi , δ ) ≤ min 8Kc′ µ c 2κ 1 d(i + 1) + ln(1/δ ) 2i+1 ˜ √ ≤ 12Kc′ µ c 2κ 1 κ 2κ −1 κ 2κ −1 2 ˜ ln(32(i + 1) /δ ) , 1 +K 2i κ 2κ −1 d(i + 1) + ln(1/δ ) = (c/2) 2i+1 . Combining this with (88) now implies κ 2κ −1 d(i + 1) + ln(1/δ ) ˆ Vi+1 ⊆ C c 2i+1 . ˆ To complete the inductive proof, it remains only to show V2⋆i+1 ⊆ Vi+1 . Toward this end, recall (i) (ii) we have shown above that on Hn ∩ Hn , V2⋆i+1 ⊆ Vi+1 upon reaching Step 9 with m = 2i+1 , and that every h1 , h2 ∈ Vi+1 at this point have erLi+1 (h1 ) − erLi+1 (h2 ) = erL⋆ (h1 ) − erL⋆ (h2 ). Consider any ˆ ˆ i+1 i+1 (i) (ii) h ∈ V2⋆i+1 , and note that any other g ∈ V2⋆i+1 has erL⋆ (g) = erL⋆ (h). Thus, on Hn ∩ Hn , i+1 i+1 erLi+1 (h) − ′min erLi+1 (h′ ) = erL⋆ (h) − ′min erL⋆ (h′ ) ˆ ˆ i+1 i+1 h ∈Vi+1 h ∈Vi+1 ≤ erL⋆ (h) − min erL⋆ (h′ ) = inf erL⋆ (g) − min erL⋆ (h′ ). (89) i+1 i+1 i+1 i+1 ⋆ ˆ h′ ∈Vi g∈V 2i+1 (i) ˆ h′ ∈Vi (ii) Lemma 58 and (86) imply that on Jn (δ ) ∩ Hn ∩ Hn , the last expression in (89) is not larger (i) ˆ ˆ than infg∈V ⋆i+1 er(g) − er( f ) + Ui+1 (Vi , δ ), and Lemma 35 implies f ∈ cl V2⋆i+1 on Hn , so that 2 infg∈V ⋆i+1 er(g) = er( f ). We therefore have 2 ˆ ˆ erLi+1 (h) − ′min erLi+1 (h′ ) ≤ Ui+1 (Vi , δ ), ˆ ˆ h ∈Vi+1 ˆ ˆ so that h ∈ Vi+1 as well. Since this holds for any h ∈ V2⋆i+1 , we have V2⋆i+1 ⊆ Vi+1 . The lemma now follows by the principle of induction. Lemma 60 There exist (C, PXY , γ )-dependent constants c∗ , c∗ ∈ [1, ∞) such that, for any ε , δ ∈ 1 2 (0, e−3 ) and integer 1 2 1 ˜ n ≥ c∗ + c∗ θ f ε κ ε κ −2 log2 , 1 2 2 εδ ∗ when running Algorithm 5 with label budget n and confidence parameter δ , on an event Jn (ε , δ ) ∗ (ε , δ )) ≥ 1 − δ , we have V ˆiˆ ⊆ C(ε ). with P(Jn ˜ df Proof Define ˜ c∗ = max 2d f +5 1 µ c1/κ r(1−γ )/6 2κ −1 d log2 d µ c1/κ 2 120 , ln 8c(i) , 1/3 ln 8c(ii) ˜ ˜ r(1−γ )/6 δ 1/3 δf f 1577 H ANNEKE and ˜ c∗ = max c∗ , 2d f +5 · 2 2κ −1 µ c1/κ r(1−γ )/6 ˜ , 2d f +15 · 1 µ c2 d 2 ˜ Fix any ε , δ ∈ (0, e−3 ) and integer n ≥ c∗ + c∗ θ f ε κ ε κ −2 log2 2 1 2 1 For each i ∈ {0, 1, . . .}, let ri = µ c κ ˜ ˜ i= 2− 1 2κ −1 di+ln(1/δ ) 2i 1 κ log2 ˜ γδ f 1 εδ log2 (4dc) . 2 . . Also define c 2dc + log2 8d log2 ε εδ . ˇ ˇ ˆ and let i = min i ∈ N : sup j≥i r j < r(1−γ )/6 . For any i ∈ i, . . . , id˜f , let ˜ (d˜ ) ˆ Qi+1 = m ∈ 2i + 1, . . . , 2i+1 : ∆4mf (Xm ,W2 , B ( f , ri )) ≥ 2γ /3 . ˜ Also define 1 2 2dc 96 ˜ ˜ θ ε κ · 2µ c2 · 8d log2 Q= · ε κ −2 . ˜f f εδ γδ (i) (ii) ˆ By Lemma 59 and Condition 1, on Jn (δ ) ∩ Hn ∩ Hn , if i ≤ id˜f , di + ln(1/δ ) ˆ Vi ⊆ C c 2i κ 2κ −1 ⊆ B ( f , ri ) . ˜ (90) (i) (ii) ˆ ˆ Lemma 59 also implies that, on Jn (δ ) ∩ Hn ∩ Hn , for i with id˜f −1 ≤ i ≤ id˜f , all of the sets Vi+1 ˆ obtained in Algorithm 5 while k = d˜f and m ∈ 2i + 1, . . . , 2i+1 satisfy V2⋆i+1 ⊆ Vi+1 ⊆ Vi . Recall that ˜f = 1 or else every m ∈ 2i + 1, . . . , 2i+1 has 4m > n. Also ˆ i1 ≥ ⌊log2 (n/2)⌋, so that we have either d (i) ˇ recall that Lemma 49 implies that when the above conditions are satisfied, and i ≥ i, on H ′ ∩ Gn , ˜f ) ˜f ) ˆ (d ˆ (d ∆4m (Xm ,W2 ,Vi+1 ) ≤ (3/2)∆4m (Xm ,W2 , B ( f , ri )), so that |Qi+1 | upper bounds the number of m ∈ ˜ i + 1, . . . , 2i+1 for which Algorithm 5 requests the label Y in Step 6 of the k = d round. Thus, ˜f 2 m (i) (ii) on Jn (δ ) ∩ Hn ∩ Hn , 2i + ˇ ˆ id˜ f ˇˆ i=max i,id˜ f −1 |Qi+1 | upper bounds the total number of label requests by Algorithm 5 while k = d˜f ; therefore, by the constraint in Step 3, we know that either this quantity ˆ i ˜ +1 ˜ is at least as big as 2−d f n , or else we have 2 d f > d˜f · 2n . In particular, on this event, if we can show that ˆ ˜ min id˜ ,i f ˜ ˜ |Qi+1 | < 2−d f n and 2i+1 ≤ d˜f · 2n , ˇ i 2+ ˇˆ i=max i,id˜ (91) f −1 ˜ ˆ then it must be true that i < id˜f . Next, we will focus on establishing this fact. ˇˆ ˆ ˜ and any m ∈ 2i + 1, . . . , 2i+1 . If d˜f = 1, Consider any i ∈ max i, id˜f −1 , . . . , min id˜f , i then ˜ ˜ ˜ ˆ (d ) ˜ ˜ P ∆4mf (Xm ,W2 , B ( f , ri )) ≥ 2γ /3 W2 = P d f S d f (B ( f , ri )) . 1578 ACTIVIZED L EARNING ˜ ˆ (d ) Otherwise, if d˜f > 1, then by Markov’s inequality and the definition of ∆4mf (·, ·, ·) from (15), ˜ ˜ 3 ˆ (d ) ˆ (d ) ˜ ˜ P ∆4mf (Xm ,W2 , B ( f , ri )) ≥ 2γ /3 W2 ≤ E ∆4mf (Xm ,W2 , B ( f , ri )) W2 2γ 1 3 = (d˜f ) 2γ M (B ( f , r )) ˜i 4m (4m)3 (d˜f ) P Ss s=1 (d˜f ) ˜ ∪ {Xm } ∈ S d f (B ( f , ri )) Ss ˜ (i) . (ii) By Lemma 39, Lemma 59, and (90), on Jn (δ ) ∩ Hn ∩ Hn , this is at most 3 1 ˜ f γ (4m)3 δ (4m)3 (d˜f ) P Ss s=1 24 1 ≤ ˜ f γ 43 23i+3 δ (d˜f ) ˜ ∪ {Xm } ∈ S d f (B ( f , ri )) Ss ˜ 43 23i+3 (d˜f ) P Ss s=1 (d˜f ) ˜ ˜ ∪ {Xm } ∈ S d f (B ( f , ri )) Ss . Note that this value is invariant to the choice of m ∈ 2i + 1, . . . , 2i+1 . By Hoeffding’s inequality, ∗ ∗ on an event Jn (i) of probability P (Jn (i)) ≥ 1 − δ /(16i2 ), this is at most ln(4i/δ ) ˜ ˜ + P d f S d f (B ( f , ri )) ˜ 43 23i+3 24 ˜ δf γ . (92) ˆ Since i ≥ i1 > log2 (n/4) and n ≥ ln(1/δ ), we have ln(4i/δ ) ≤ 2−i 43 23i+3 ln(4 log2 (n/4)/δ ) ≤ 2−i 128n ln(n/δ ) ≤ 2−i . 128n Thus, (92) is at most 24 ˜ ˜ 2−i + P d f S d f (B ( f , ri )) ˜ ˜f γ δ . 1 (i) (ii) ∗ ˜ In either case (d˜f = 1 or d˜f > 1), by definition of θ f ε κ , on Jn (δ ) ∩ Hn ∩ Hn ∩ Jn (i), ∀m ∈ 2i + 1, . . . , 2i+1 we have ˜ 1 1 24 ˜ ˆ (d ) 2−i + θ f ε κ · max ri , ε κ ˜ ˜ P ∆4mf (Xm ,W2 , B ( f , ri )) ≥ 2γ /3 W2 ≤ ˜ δf γ . (93) ˜ ˆ (d ) Furthermore, the ½[2γ /3,∞) ∆4mf (Xm ,W2 , B ( f , ri )) indicators are conditionally independent given ˜ ˜ W2 , so that we may bound P |Qi+1 | > Q W2 via a Chernoff bound. Toward this end, note that on (i) (ii) ∗ Jn (δ ) ∩ Hn ∩ Hn ∩ Jn (i), (93) implies 2i+1 E |Qi+1 | W2 = ≤ 2i · m=2i +1 ˜ ˆ (d ) ˜ P ∆4mf (Xm ,W2 , B ( f , ri )) ≥ 2γ /3 W2 1 1 24 ˜ 2−i + θ f ε κ · max ri , ε κ ˜ ˜ δf γ ≤ 1 24 ˜ 1 ˜ 1 + θ f ε κ · max 2i ri , 2i ε κ ˜ ˜ δf γ 1579 . (94) H ANNEKE Note that 1 2i ri = µ c κ (di + ln(1/δ )) 2κ −1 · 2i(1− 2κ −1 ) ˜ 1 ≤ µc ˜ 1 1 κ Then since 2−i 2κ −1 ≤ most ˜ d i + ln(1/δ ) ε c 1 κ 1 2κ −1 1 ·2 · 8d log2 2dc εδ 1 24 ˜ 1 ˜ 1 + θ f ε κ · µ · 2i ε κ ˜f γδ ≤ ˜ i(1− 2κ1 ) −1 − 2κ1 −1 ≤ µc 2dc 8d log2 εδ 1 κ 1 2κ −1 · 2i(1− 2κ −1 ) . ˜ , we have that the rightmost expression in (94) is at 1 2 2dc ˜ · ε κ −2 1 + θ f ε κ · 2µ c2 · 8d log2 εδ 24 ˜ γδ f 1 (i) ˜ ≤ Q/2. (ii) ∗ Therefore, a Chernoff bound implies that on Jn (δ ) ∩ Hn ∩ Hn ∩ Jn (i), we have 2dc εδ ˜ ˜ P |Qi+1 | > Q W2 ≤ exp −Q/6 ≤ exp −8 log2 ≤ exp − log2 48 log2 (2dc/εδ ) δ ˜ ≤ δ /(8i). Combined with the law of total probability and a union bound over i values, this implies there exists (i) (ii) ∗ an event Jn (ε , δ ) ⊆ Jn (δ ) ∩ Hn ∩ Hn with ˜ i (i) (ii) ∗ P Jn (δ ) ∩ Hn ∩ Hn \ Jn (ε , δ ) ≤ ˇ i=i ˜ δ /(16i2 ) + δ /(8i) ≤ δ /4, ˜ has |Qi+1 | ≤ Q. ˜ ˜ ˇ We have chosen c∗ and c∗ large enough that 2i+1 < d˜f · 2n and 2i < 2−d f −2 n. In particular, this 2 1 ∗ (ε , δ ), means that on Jn ˇˆ ˆ ˜ on which every i ∈ max i, id˜f −1 , . . . , min id˜f , i ˜ˆ min i,id˜ f ˜ ˇ i ˜˜ |Qi+1 | < 2−d f −2 n + iQ. 2+ ˇˆ i=max i,id˜ f −1 ˜ Furthermore, since i ≤ 3 log2 4dc , we have εδ 13 2 1 2 ˜ ˜ ˜ 2 µ c d θ f ε κ · ε κ −2 · log2 4dc iQ ≤ 2 ˜f εδ γδ ≤ 1 2 1 213 µ c2 d log2 (4dc) ˜ ˜ 2 θ f ε κ · ε κ −2 · log2 ≤ 2−d f −2 n. 2 ˜f εδ γδ ∗ ˜ ˆ Combining the above, we have that (91) is satisfied on Jn (ε , δ ), so that id˜f > i. Combined with ∗ (ε , δ ), Lemma 59, this implies that on Jn ˆ Viˆ ˜ df ˜ d i + ln(1/δ ) ˆ ⊆ Vi˜ ⊆ C c 2i˜ 1580 κ 2κ −1 , ACTIVIZED L EARNING ˜ and by definition of i we have ˜ d i + ln(1/δ ) c 2i˜ κ 2κ −1 2dc ≤ c 8d log2 εδ 2dc ≤ c 8d log2 εδ κ 2κ −1 κ 2κ −1 ˜ κ · 2−i 2κ −1 2dc · (ε /c) · 8d log2 εδ − 2κκ −1 = ε, ˆ so that Viˆ ˜ ⊆ C(ε ). df ∗ Finally, to prove the stated bound on P(Jn (ε , δ )), by a union bound we have (i) ∗ 1 − P (Jn (ε , δ )) ≤ (1 − P(Jn (δ ))) + 1 − P Hn (i) (i) (ii) + P Hn \ Hn (ii) ∗ + P Jn (δ ) ∩ Hn ∩ Hn \ Jn (ε , δ ) 1/3 ˜ ˜ ≤ 3δ /4 + c(i) · exp −n3 δ f /8 + c(ii) · exp −nδ f /120 ≤ δ . We are now ready for the proof of Lemma 26. Proof [Lemma 26] First, note that because we break ties in the argmax of Step 7 in favor of a y value ˆ with Vik +1 [(Xm , y)] = ∅, if Vik +1 = ∅ before Step 8, then this remains true after Step 8. Furthermore, ˆ ˆ the Uik +1 estimator is nonnegative, and thus the update in Step 10 never removes from Vik +1 the minimizer of erLi +1 (h) among h ∈ Vik +1 . Therefore, by induction we have Vik = ∅ at all times in ˆ k ˆ ˆ Algorithm 5. In particular, Viˆd+1 +1 = ∅ so that the return classifier h exists. Also, by Lemma 60, for ∗ (ε , δ ), running Algorithm 5 with label budget n and confidence parameter n as in Lemma 60, on Jn ∗ ˆ δ results in Viˆ ˜ ⊆ C(ε ). Combining these two facts implies that for such a value of n, on Jn (ε , δ ), df ˆ ˆ ˆ ˆ h ∈ Viˆd+1 +1 ⊆ Viˆ ˜ ⊆ C(ε ), so that er h ≤ ν + ε . df E.3 The Misspecified Model Case Here we present a proof of Theorem 28, including a specification of the method A′ from the theorem a statement. Proof [Theorem 28] Consider a weakly universally consistent passive learning algorithm Au (Devroye, Gy¨ rfi, and Lugosi, 1996). Such a method must exist in our setting; for instance, Hoeffding’s o inequality and a union bound imply that it suffices to take Au (L) = argmin½± erL (½±i ) + ln(4i |L|) , B 2|L| Bi where {B1 , B2 , . . .} is a countable algebra that generates FX . Then Au achieves a label complexity Λu such that for any distribution PXY on X × {−1, +1}, ∀ε ∈ (0, 1), Λu (ε + ν ∗ (PXY ), PXY ) < ∞. In particular, if ν ∗ (PXY ) < ν (C; PXY ), then we have Λu ((ν ∗ (PXY ) + ν (C; PXY ))/2, PXY ) < ∞. Fix any n ∈ N and describe the execution of A′ (n) as follows. In a preprocessing step, witha hold the first mun = n − ⌊n/2⌋ − ⌊n/3⌋ ≥ n/6 examples {X1 , . . . , Xmun } and request their labels {Y1 , . . . ,Ymun }. Run Aa (⌊n/2⌋) on the remainder of the sequence {Xmun +1 , Xmun +2 , . . .} (i.e., shift 2 1581 H ANNEKE any index references in the algorithm by mun ), and let ha denote the classifier it returns. Also request the labels Ymun +1 , . . .Ymun +⌊n/3⌋ , and let hu = Au (Xmun +1 ,Ymun +1 ), . . . , (Xmun +⌊n/3⌋ ,Ymun +⌊n/3⌋ ) . ˆ ˆ If ermun (ha ) − ermun (hu ) > n−1/3 , return h = hu ; otherwise, return h = ha . This method achieves the stated result, for the following reasons. First, let us examine the final step of this algorithm. By Hoeffding’s inequality, with probability at least 1 − 2 · exp −n1/3 /12 , |(ermun (ha ) − ermun (hu )) − (er(ha ) − er(hu ))| ≤ n−1/3 . ˆ When this is the case, a triangle inequality implies er(h) ≤ min{er(ha ), er(hu ) + 2n−1/3 }. If PXY satisfies the benign noise case, then for any n ≥ 2Λa (ε /2 + ν (C; PXY ), PXY ), ˆ we have E[er(ha )] ≤ ν (C; PXY ) + ε /2, so E[er(h)] ≤ ν (C; PXY ) + ε /2 + 2 · exp{−n1/3 /12}, which 3 ln3 (4/ε ). So in this case, we can take λ (ε ) = 123 ln3 (4/ε ) . is at most ν (C; PXY ) + ε if n ≥ 12 On the other hand, if PXY is not in the benign noise case (i.e., the misspecified model case), then for any n ≥ 3Λu ((ν ∗ (PXY ) + ν (C; PXY ))/2, PXY ), E [er(hu )] ≤ (ν ∗ (PXY ) + ν (C; PXY ))/2, so that ˆ E[er(h)] ≤ E[er(hu )] + 2n−1/3 + 2 · exp{−n1/3 /12} ≤ (ν ∗ (PXY ) + ν (C; PXY ))/2 + 2n−1/3 + 2 · exp{−n1/3 /12}. 2 Again, this is at most ν (C; PXY ) + ε if n ≥ max 123 ln3 ε , 64(ν (C; PXY ) − ν ∗ (PXY ))−3 . So in this case, we can take 2 ν ∗ (PXY ) + ν (C; PXY ) 64 λ (ε ) = max 123 ln3 , 3Λu , PXY , ε 2 (ν (C; PXY ) − ν ∗ (PXY ))3 . In either case, we have λ (ε ) ∈ Polylog(1/ε ). References N. Abe and H. Mamitsuka. Query learning strategies using boosting and bagging. In Proceedings of the 15th International Conference on Machine Learning, 1998. M. Alekhnovich, M. Braverman, V. Feldman, A. Klivans, and T. Pitassi. Learnability and automatizability. In Proceedings of the 45th Foundations of Computer Science, 2004. K. Alexander. Probability inequalities for empirical processes and a law of the iterated logarithm. The Annals of Probability, 4:1041–1067, 1984. M. Anthony and P. L. Bartlett. Neural Network Learning: Theoretical Foundations. Cambridge University Press, 1999. 1582 ACTIVIZED L EARNING A. Antos and G. Lugosi. Strong minimax lower bounds for learning. Machine Learning, 30:31–56, 1998. R. B. Ash and C. A. Dol´ ans-Dade. Probability & Measure Theory. Academic Press, 2000. e M.-F. Balcan, A. Beygelzimer, and J. Langford. Agnostic active learning. In Proceedings of the 23rd International Conference on Machine Learning, 2006a. M.-F. Balcan, A. Blum, and S. Vempala. Kernels as features: On kernels, margins, and lowdimensional mappings. Machine Learning Journal, 65(1):79–94, 2006b. M.-F. Balcan, A. Broder, and T. Zhang. Margin based active learning. In Proceedings of the 20th Conference on Learning Theory, 2007. M.-F. Balcan, A. Beygelzimer, and J. Langford. Agnostic active learning. Journal of Computer and System Sciences, 75(1):78–89, 2009. M.-F. Balcan, S. Hanneke, and J. Wortman Vaughan. The true sample complexity of active learning. Machine Learning, 80(2–3):111–139, 2010. J. Baldridge and A. Palmer. How well does active learning actually work? Time-based evaluation of cost-reduction strategies for language documentation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2009. Z. Bar-Yossef. Sampling lower bounds via information theory. In Proceedings of the 35th Annual ACM Symposium on the Theory of Computing, 2003. P. L. Bartlett, M. I. Jordan, and J. D. McAuliffe. Convexity, classification, and risk bounds. Journal of the American Statistical Association, 101(473):138–156, 2006. A. Beygelzimer, S. Dasgupta, and J. Langford. Importance weighted active learning. In Proceedings of the International Conference on Machine Learning, 2009. A. Beygelzimer, D. Hsu, J. Langford, and T. Zhang. Agnostic active learning without constraints. In Advances in Neural Information Processing Systems 23, 2010. A. Blumer, A. Ehrenfeucht, D. Haussler, and M. Warmuth. Learnability and the VapnikChervonenkis dimension. Journal of the Association for Computing Machinery, 36(4):929–965, 1989. F. Bunea, A. B. Tsybakov, and M. Wegkamp. Sparsity oracle inequalities for the lasso. Electronic Journal of Statistics, 1:169–194, 2009. C. Campbell, N. Cristianini, and A. Smola. Query learning with large margin classifiers. In Proceedings of the 17th International Conference on Machine Learning, 2000. R. Castro and R. D. Nowak. Minimax bounds for active learning. IEEE Transactions on Information Theory, 54(5):2339–2353, 2008. D. Cohn, L. Atlas, and R. Ladner. Improving generalization with active learning. Machine Learning, 15(2):201–221, 1994. 1583 H ANNEKE S. Dasgupta. Coarse sample complexity bounds for active learning. In Advances in Neural Information Processing Systems 18, 2005. S. Dasgupta, A. T. Kalai, and C. Monteleoni. Analysis of perceptron-based active learning. In Proceedings of the 18th Conference on Learning Theory, 2005. S. Dasgupta, D. Hsu, and C. Monteleoni. A general agnostic active learning algorithm. In Advances in Neural Information Processing Systems 20, 2007. S. Dasgupta, A. T. Kalai, and C. Monteleoni. Analysis of perceptron-based active learning. Journal of Machine Learning Research, 10:281–299, 2009. O. Dekel, C. Gentile, and K. Sridharan. Robust selective sampling from single and multiple teachers. In Proceedings of the 23rd Conference on Learning Theory, 2010. L. Devroye, L. Gy¨ rfi, and G. Lugosi. A Probabilistic Theory of Pattern Recognition. Springero Verlag New York, Inc., 1996. R. M. Dudley. Real Analysis and Probability. Cambridge University Press, 2002. Y. Freund, H. S. Seung, E. Shamir, and N. Tishby. Selective sampling using the query by committee algorithm. Machine Learning, 28:133–168, 1997. E. Friedman. Active learning for smooth problems. In Proceedings of the 22nd Conference on Learning Theory, 2009. R. Gangadharaiah, R. D. Brown, and J. Carbonell. Active learning in example-based machine translation. In Proceedings of the 17th Nordic Conference on Computational Linguistics, 2009. E. Gin´ and V. Koltchinskii. Concentration inequalities and asymptotic results for ratio type empire ical processes. The Annals of Probability, 34(3):1143–1216, 2006. S. A. Goldman and M. J. Kearns. On the complexity of teaching. Journal of Computer and System Sciences, 50:20–31, 1995. S. Hanneke. Teaching dimension and the complexity of active learning. In Proceedings of the 20th Conference on Learning Theory, 2007a. S. Hanneke. A bound on the label complexity of agnostic active learning. In Proceedings of the 24th International Conference on Machine Learning, 2007b. S. Hanneke. Adaptive rates of convergence in active learning. In Proceedings of the 22nd Conference on Learning Theory, 2009a. S. Hanneke. Theoretical Foundations of Active Learning. PhD thesis, Machine Learning Department, School of Computer Science, Carnegie Mellon University, 2009b. S. Hanneke. Rates of convergence in active learning. The Annals of Statistics, 39(1):333–361, 2011. S. Har-Peled, D. Roth, and D. Zimak. Maximum margin coresets for active and noise tolerant learning. In Proceedings of the 20th International Joint Conference on Artificial Intelligence, 2007. 1584 ACTIVIZED L EARNING D. Haussler. Decision theoretic generalizations of the PAC model for neural net and other learning applications. Information and Computation, 100:78–150, 1992. D. Haussler, N. Littlestone, and M. Warmuth. Predicting {0, 1}-functions on randomly drawn points. Information and Computation, 115:248–292, 1994. T. Heged¨ s. Generalized teaching dimension and the query complexity of learning. In Proceedings u of the 8th Conference on Computational Learning Theory, 1995. L. Hellerstein, K. Pillaipakkamnatt, V. Raghavan, and D. Wilkins. How many queries are needed to learn? Journal of the Association for Computing Machinery, 43(5):840–862, 1996. D. Helmbold, R. Sloan, and M. Warmuth. Learning nested differences of intersection-closed concept classes. Machine Learning, 5:165–196, 1990. S. C. H. Hoi, R. Jin, J. Zhu, and M. R. Lyu. Batch mode active learning and its application to medical image classification. In Proceedings of the 23rd International Conference on Machine Learning, 2006. M. K¨ ari¨ inen. Active learning in the non-realizable case. In Proceedings of the 17th International a¨ a Conference on Algorithmic Learning Theory, 2006. N. Karmarkar. A new polynomial-time algorithm for linear programming. Combinatorica, 4:373– 395, 1984. M. J. Kearns and U. Vazirani. An Introduction to Computational Learning Theory. The MIT Press, 1994. M. J. Kearns, R. E. Schapire, and L. M. Sellie. Toward efficient agnostic learning. Machine Learning, 17:115–141, 1994. L. G. Khachiyan. A polynomial algorithm in linear programming. Soviet Mathematics Doklady, 20:191–194, 1979. V. Koltchinskii. Local Rademacher complexities and oracle inequalities in risk minimization. The Annals of Statistics, 34(6):2593–2656, 2006. V. Koltchinskii. Rademacher complexities and bounding the excess risk in active learning. Journal of Machine Learning Research, 11:2457–2485, 2010. V. Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems. In ´ ´e Ecole d’Et´ de Probabilit´ s de Saint-Flour XXXVIII-2008. Lecture Notes in Mathematics, 2033, e Springer, 2011. S. Li. Concise formulas for the area and volume of a hyperspherical cap. Asian Journal of Mathematics and Statistics, 4(1):66–70, 2011. M. Lindenbaum, S. Markovitch, and D. Rusakov. Selective sampling for nearest neighbor classifiers. Machine Learning, 54:125–152, 2004. 1585 H ANNEKE T. Luo, K. Kramer, D. B. Goldgof, L. O. Hall, S. Samson, A. Remsen, and T. Hopkins. Active learning to recognize multiple types of plankton. Journal of Machine Learning Research, 6: 589–613, 2005. S. Mahalanabis. A note on active learning for smooth problems. arXiv:1103.3095, 2011. E. Mammen and A. B. Tsybakov. Smooth discrimination analysis. The Annals of Statistics, 27: 1808–1829, 1999. ´ e e P. Massart and E. N´ d´ lec. Risk bounds for statistical learning. The Annals of Statistics, 34(5): 2326–2366, 2006. A. McCallum and K. Nigam. Employing EM in pool-based active learning for text classification. In Proceedings of the 15th International Conference on Machine Learning, 1998. P. Mitra, C. A. Murthy, and S. K. Pal. A probabilistic active support vector learning algorithm. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(3):413–418, 2004. J. R. Munkres. Topology. Prentice Hall, Inc., 2nd edition, 2000. I. Muslea, S. Minton, and C. A. Knoblock. Active + semi-supervised learning = robust multi-view learning. In Proceedings of the 19th International Conference on Machine Learning, 2002. R. D. Nowak. Generalized binary search. In Proceedings of the 46th Annual Allerton Conference on Communication, Control, and Computing, 2008. L. Pitt and L. G. Valiant. Computational limitations on learning from examples. Journal of the Association for Computing Machinery, 35(4):965–984, 1988. J. Poland and M. Hutter. MDL convergence speed for Bernoulli sequences. Statistics and Computing, 16:161–175, 2006. G. V. Rocha, X. Wang, and B. Yu. Asymptotic distribution and sparsistency for l1-penalized parametric M-estimators with applications to linear SVM and logistic regression. arXiv:0908.1940v1, 2009. D. Roth and K. Small. Margin-based active learning for structured output spaces. In European Conference on Machine Learning, 2006. N. Roy and A. McCallum. Toward optimal active learning through sampling estimation of error reduction. In Proceedings of the 18th International Conference on Machine Learning, 2001. A. I. Schein and L. H. Ungar. Active learning for logistic regression: An evaluation. Machine Learning, 68(3):235–265, 2007. G. Schohn and D. Cohn. Less is more: Active learning with support vector machines. In Proceedings of the 17th International Conference on Machine Learning, 2000. B. Settles. Active learning literature survey. http://active-learning.net, 2010. S. M. Srivastava. A Course on Borel Sets. Springer-Verlag, 1998. 1586 ACTIVIZED L EARNING S. Tong and D. Koller. Support vector machine active learning with applications to text classification. Journal of Machine Learning Research, 2, 2001. A. B. Tsybakov. Optimal aggregation of classifiers in statistical learning. The Annals of Statistics, 32(1):135–166, 2004. L. G. Valiant. A theory of the learnable. Communications of the Association for Computing Machinery, 27(11):1134–1142, 1984. A. W. van der Vaart and J. A. Wellner. Weak Convergence and Empirical Processes. Springer, 1996. V. Vapnik. Estimation of Dependencies Based on Empirical Data. Springer-Verlag, New York, 1982. V. Vapnik. Statistical Learning Theory. John Wiley & Sons, Inc., 1998. V. Vapnik and A. Chervonenkis. On the uniform convergence of relative frequencies of events to their probabilities. Theory of Probability and its Applications, 16:264–280, 1971. A. Wald. Sequential tests of statistical hypotheses. The Annals of Mathematical Statistics, 16(2): 117–186, 1945. L. Wang. Sufficient conditions for agnostic active learnable. In Advances in Neural Information Processing Systems 22, 2009. L. Wang. Smoothness, disagreement coefficient, and the label complexity of agnostic active learning. Journal of Machine Learning Research, 12:2269–2292, 2011. L. Wang and X. Shen. On L1-norm multiclass support vector machines. Journal of the American Statistical Association, 102(478):583–594, 2007. L. Yang, S. Hanneke, and J. Carbonell. The sample complexity of self-verifying Bayesian active learning. In Proceedings of the 14th International Conference on Artificial Intelligence and Statistics, 2011. 1587
15 jmlr-2012-Algebraic Geometric Comparison of Probability Distributions
Author: Franz J. Király, Paul von Bünau, Frank C. Meinecke, Duncan A.J. Blythe, Klaus-Robert Müller
Abstract: We propose a novel algebraic algorithmic framework for dealing with probability distributions represented by their cumulants such as the mean and covariance matrix. As an example, we consider the unsupervised learning problem of finding the subspace on which several probability distributions agree. Instead of minimizing an objective function involving the estimated cumulants, we show that by treating the cumulants as elements of the polynomial ring we can directly solve the problem, at a lower computational cost and with higher accuracy. Moreover, the algebraic viewpoint on probability distributions allows us to invoke the theory of algebraic geometry, which we demonstrate in a compact proof for an identifiability criterion. Keywords: computational algebraic geometry, approximate algebra, unsupervised Learning
16 jmlr-2012-Algorithms for Learning Kernels Based on Centered Alignment
Author: Corinna Cortes, Mehryar Mohri, Afshin Rostamizadeh
Abstract: This paper presents new and effective algorithms for learning kernels. In particular, as shown by our empirical results, these algorithms consistently outperform the so-called uniform combination solution that has proven to be difficult to improve upon in the past, as well as other algorithms for learning kernels based on convex combinations of base kernels in both classification and regression. Our algorithms are based on the notion of centered alignment which is used as a similarity measure between kernels or kernel matrices. We present a number of novel algorithmic, theoretical, and empirical results for learning kernels based on our notion of centered alignment. In particular, we describe efficient algorithms for learning a maximum alignment kernel by showing that the problem can be reduced to a simple QP and discuss a one-stage algorithm for learning both a kernel and a hypothesis based on that kernel using an alignment-based regularization. Our theoretical results include a novel concentration bound for centered alignment between kernel matrices, the proof of the existence of effective predictors for kernels with high alignment, both for classification and for regression, and the proof of stability-based generalization bounds for a broad family of algorithms for learning kernels based on centered alignment. We also report the results of experiments with our centered alignment-based algorithms in both classification and regression. Keywords: kernel methods, learning kernels, feature selection
Author: Nir Ailon
Abstract: Given a set V of n elements we wish to linearly order them given pairwise preference labels which may be non-transitive (due to irrationality or arbitrary noise). The goal is to linearly order the elements while disagreeing with as few pairwise preference labels as possible. Our performance is measured by two parameters: The number of disagreements (loss) and the query complexity (number of pairwise preference labels). Our algorithm adaptively queries at most O(ε−6 n log5 n) preference labels for a regret of ε times the optimal loss. As a function of n, this is asymptotically better than standard (non-adaptive) learning bounds achievable for the same problem. Our main result takes us a step closer toward settling an open problem posed by learning-torank (from pairwise information) theoreticians and practitioners: What is a provably correct way to sample preference labels? To further show the power and practicality of our solution, we analyze a typical test case in which a large margin linear relaxation is used for efficiently solving the simpler learning problems in our decomposition. Keywords: statistical learning theory, active learning, ranking, pairwise ranking, preferences
18 jmlr-2012-An Improved GLMNET for L1-regularized Logistic Regression
Author: Guo-Xun Yuan, Chia-Hua Ho, Chih-Jen Lin
Abstract: Recently, Yuan et al. (2010) conducted a comprehensive comparison on software for L1-regularized classification. They concluded that a carefully designed coordinate descent implementation CDN is the fastest among state-of-the-art solvers. In this paper, we point out that CDN is less competitive on loss functions that are expensive to compute. In particular, CDN for logistic regression is much slower than CDN for SVM because the logistic loss involves expensive exp/log operations. In optimization, Newton methods are known to have fewer iterations although each iteration costs more. Because solving the Newton sub-problem is independent of the loss calculation, this type of methods may surpass CDN under some circumstances. In L1-regularized classification, GLMNET by Friedman et al. is already a Newton-type method, but experiments in Yuan et al. (2010) indicated that the existing GLMNET implementation may face difficulties for some largescale problems. In this paper, we propose an improved GLMNET to address some theoretical and implementation issues. In particular, as a Newton-type method, GLMNET achieves fast local convergence, but may fail to quickly obtain a useful solution. By a careful design to adjust the effort for each iteration, our method is efficient for both loosely or strictly solving the optimization problem. Experiments demonstrate that our improved GLMNET is more efficient than CDN for L1-regularized logistic regression. Keywords: L1 regularization, linear classification, optimization methods, logistic regression, support vector machines
19 jmlr-2012-An Introduction to Artificial Prediction Markets for Classification
Author: Adrian Barbu, Nathan Lay
Abstract: Prediction markets are used in real life to predict outcomes of interest such as presidential elections. This paper presents a mathematical theory of artificial prediction markets for supervised learning of conditional probability estimators. The artificial prediction market is a novel method for fusing the prediction information of features or trained classifiers, where the fusion result is the contract price on the possible outcomes. The market can be trained online by updating the participants’ budgets using training examples. Inspired by the real prediction markets, the equations that govern the market are derived from simple and reasonable assumptions. Efficient numerical algorithms are presented for solving these equations. The obtained artificial prediction market is shown to be a maximum likelihood estimator. It generalizes linear aggregation, existent in boosting and random forest, as well as logistic regression and some kernel methods. Furthermore, the market mechanism allows the aggregation of specialized classifiers that participate only on specific instances. Experimental comparisons show that the artificial prediction markets often outperform random forest and implicit online learning on synthetic data and real UCI data sets. Moreover, an extensive evaluation for pelvic and abdominal lymph node detection in CT data shows that the prediction market improves adaboost’s detection rate from 79.6% to 81.2% at 3 false positives/volume. Keywords: online learning, ensemble methods, supervised learning, random forest, implicit online learning
20 jmlr-2012-Analysis of a Random Forests Model
Author: Gérard Biau
Abstract: Random forests are a scheme proposed by Leo Breiman in the 2000’s for building a predictor ensemble with a set of decision trees that grow in randomly selected subspaces of data. Despite growing interest and practical use, there has been little exploration of the statistical properties of random forests, and little is known about the mathematical forces driving the algorithm. In this paper, we offer an in-depth analysis of a random forests model suggested by Breiman (2004), which is very close to the original algorithm. We show in particular that the procedure is consistent and adapts to sparsity, in the sense that its rate of convergence depends only on the number of strong features and not on how many noise variables are present. Keywords: random forests, randomization, sparsity, dimension reduction, consistency, rate of convergence
21 jmlr-2012-Bayesian Mixed-Effects Inference on Classification Performance in Hierarchical Data Sets
22 jmlr-2012-Bounding the Probability of Error for High Precision Optical Character Recognition
24 jmlr-2012-Causal Bounds and Observable Constraints for Non-deterministic Models
26 jmlr-2012-Coherence Functions with Applications in Large-Margin Classification Methods
28 jmlr-2012-Confidence-Weighted Linear Classification for Text Categorization
29 jmlr-2012-Consistent Model Selection Criteria on High Dimensions
30 jmlr-2012-DARWIN: A Framework for Machine Learning and Computer Vision Research and Development
31 jmlr-2012-DEAP: Evolutionary Algorithms Made Easy
32 jmlr-2012-Discriminative Hierarchical Part-based Models for Human Parsing and Action Recognition
33 jmlr-2012-Distance Metric Learning with Eigenvalue Optimization
34 jmlr-2012-Dynamic Policy Programming
35 jmlr-2012-EP-GIG Priors and Applications in Bayesian Sparse Learning
36 jmlr-2012-Efficient Methods for Robust Classification Under Uncertainty in Kernel Matrices
37 jmlr-2012-Eliminating Spammers and Ranking Annotators for Crowdsourced Labeling Tasks
38 jmlr-2012-Entropy Search for Information-Efficient Global Optimization
40 jmlr-2012-Exact Covariance Thresholding into Connected Components for Large-Scale Graphical Lasso
41 jmlr-2012-Exploration in Relational Domains for Model-based Reinforcement Learning
42 jmlr-2012-Facilitating Score and Causal Inference Trees for Large Observational Studies
43 jmlr-2012-Fast Approximation of Matrix Coherence and Statistical Leverage
44 jmlr-2012-Feature Selection via Dependence Maximization
46 jmlr-2012-Finite-Sample Analysis of Least-Squares Policy Iteration
47 jmlr-2012-GPLP: A Local and Parallel Computation Toolbox for Gaussian Process Regression
49 jmlr-2012-Hope and Fear for Discriminative Training of Statistical Translation Models
50 jmlr-2012-Human Gesture Recognition on Product Manifolds
51 jmlr-2012-Integrating a Partial Model into Model Free Reinforcement Learning
52 jmlr-2012-Iterative Reweighted Algorithms for Matrix Rank Minimization
54 jmlr-2012-Large-scale Linear Support Vector Regression
55 jmlr-2012-Learning Algorithms for the Classification Restricted Boltzmann Machine
56 jmlr-2012-Learning Linear Cyclic Causal Models with Latent Variables
57 jmlr-2012-Learning Symbolic Representations of Hybrid Dynamical Systems
58 jmlr-2012-Linear Fitted-Q Iteration with Multiple Reward Functions
59 jmlr-2012-Linear Regression With Random Projections
60 jmlr-2012-Local and Global Scaling Reduce Hubs in Space
61 jmlr-2012-ML-Flex: A Flexible Toolbox for Performing Classification Analyses In Parallel
62 jmlr-2012-MULTIBOOST: A Multi-purpose Boosting Package
63 jmlr-2012-Mal-ID: Automatic Malware Detection Using Common Segment Analysis and Meta-Features
64 jmlr-2012-Manifold Identification in Dual Averaging for Regularized Stochastic Online Learning
65 jmlr-2012-MedLDA: Maximum Margin Supervised Topic Models
66 jmlr-2012-Metric and Kernel Learning Using a Linear Transformation
68 jmlr-2012-Minimax Manifold Estimation
69 jmlr-2012-Mixability is Bayes Risk Curvature Relative to Log Loss
70 jmlr-2012-Multi-Assignment Clustering for Boolean Data
71 jmlr-2012-Multi-Instance Learning with Any Hypothesis Class
72 jmlr-2012-Multi-Target Regression with Rule Ensembles
73 jmlr-2012-Multi-task Regression using Minimal Penalties
74 jmlr-2012-Multi Kernel Learning with Online-Batch Optimization
75 jmlr-2012-NIMFA : A Python Library for Nonnegative Matrix Factorization
77 jmlr-2012-Non-Sparse Multiple Kernel Fisher Discriminant Analysis
78 jmlr-2012-Nonparametric Guidance of Autoencoder Representations using Label Information
79 jmlr-2012-Oger: Modular Learning Architectures For Large-Scale Sequential Processing
80 jmlr-2012-On Ranking and Generalization Bounds
81 jmlr-2012-On the Convergence Rate oflp-Norm Multiple Kernel Learning
82 jmlr-2012-On the Necessity of Irrelevant Variables
83 jmlr-2012-Online Learning in the Embedded Manifold of Low-rank Matrices
84 jmlr-2012-Online Submodular Minimization
85 jmlr-2012-Optimal Distributed Online Prediction Using Mini-Batches
86 jmlr-2012-Optimistic Bayesian Sampling in Contextual-Bandit Problems
87 jmlr-2012-PAC-Bayes Bounds with Data Dependent Priors
88 jmlr-2012-PREA: Personalized Recommendation Algorithms Toolkit
89 jmlr-2012-Pairwise Support Vector Machines and their Application to Large Scale Problems
90 jmlr-2012-Pattern for Python
91 jmlr-2012-Plug-in Approach to Active Learning
92 jmlr-2012-Positive Semidefinite Metric Learning Using Boosting-like Algorithms
93 jmlr-2012-Quantum Set Intersection and its Application to Associative Memory
94 jmlr-2012-Query Strategies for Evading Convex-Inducing Classifiers
95 jmlr-2012-Random Search for Hyper-Parameter Optimization
96 jmlr-2012-Refinement of Operator-valued Reproducing Kernels
97 jmlr-2012-Regularization Techniques for Learning with Matrices
98 jmlr-2012-Regularized Bundle Methods for Convex and Non-Convex Risks
99 jmlr-2012-Restricted Strong Convexity and Weighted Matrix Completion: Optimal Bounds with Noise
100 jmlr-2012-Robust Kernel Density Estimation
101 jmlr-2012-SVDFeature: A Toolkit for Feature-based Collaborative Filtering
102 jmlr-2012-Sally: A Tool for Embedding Strings in Vector Spaces
103 jmlr-2012-Sampling Methods for the Nyström Method
104 jmlr-2012-Security Analysis of Online Centroid Anomaly Detection
105 jmlr-2012-Selective Sampling and Active Learning from Single and Multiple Teachers
106 jmlr-2012-Sign Language Recognition using Sub-Units
107 jmlr-2012-Smoothing Multivariate Performance Measures
108 jmlr-2012-Sparse and Unique Nonnegative Matrix Factorization Through Data Preprocessing
109 jmlr-2012-Stability of Density-Based Clustering
110 jmlr-2012-Static Prediction Games for Adversarial Learning Problems
111 jmlr-2012-Structured Sparsity and Generalization
112 jmlr-2012-Structured Sparsity via Alternating Direction Methods
113 jmlr-2012-The huge Package for High-dimensional Undirected Graph Estimation in R
114 jmlr-2012-Towards Integrative Causal Analysis of Heterogeneous Data Sets and Studies
115 jmlr-2012-Trading Regret for Efficiency: Online Convex Optimization with Long Term Constraints
116 jmlr-2012-Transfer in Reinforcement Learning via Shared Features
117 jmlr-2012-Variable Selection in High-dimensional Varying-coefficient Models with Global Optimality
118 jmlr-2012-Variational Multinomial Logit Gaussian Process
119 jmlr-2012-glm-ie: Generalised Linear Models Inference & Estimation Toolbox