nips nips2008 nips2008-111 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Zaïd Harchaoui, Eric Moulines, Francis R. Bach
Abstract: We introduce a kernel-based method for change-point analysis within a sequence of temporal observations. Change-point analysis of an unlabelled sample of observations consists in, first, testing whether a change in the distribution occurs within the sample, and second, if a change occurs, estimating the change-point instant after which the distribution of the observations switches from one distribution to another different distribution. We propose a test statistic based upon the maximum kernel Fisher discriminant ratio as a measure of homogeneity between segments. We derive its limiting distribution under the null hypothesis (no change occurs), and establish the consistency under the alternative hypothesis (a change occurs). This allows to build a statistical hypothesis testing procedure for testing the presence of a change-point, with a prescribed false-alarm probability and detection probability tending to one in the large-sample setting. If a change actually occurs, the test statistic also yields an estimator of the change-point location. Promising experimental results in temporal segmentation of mental tasks from BCI data and pop song indexation are presented. 1
Reference: text
sentIndex sentText sentNum sentScore
1 Kernel Change-point Analysis Za¨d Harchaoui ı LTCI, TELECOM ParisTech and CNRS 46, rue Barrault, 75634 Paris cedex 13, France zaid. [sent-1, score-0.282]
2 fr Francis Bach Willow Project, INRIA-ENS 45, rue d’Ulm, 75230 Paris, France francis. [sent-3, score-0.161]
3 org ´ Eric Moulines LTCI, TELECOM ParisTech and CNRS 46, rue Barrault, 75634 Paris cedex 13, France eric. [sent-5, score-0.282]
4 fr Abstract We introduce a kernel-based method for change-point analysis within a sequence of temporal observations. [sent-7, score-0.202]
5 We propose a test statistic based upon the maximum kernel Fisher discriminant ratio as a measure of homogeneity between segments. [sent-9, score-0.469]
6 We derive its limiting distribution under the null hypothesis (no change occurs), and establish the consistency under the alternative hypothesis (a change occurs). [sent-10, score-0.837]
7 This allows to build a statistical hypothesis testing procedure for testing the presence of a change-point, with a prescribed false-alarm probability and detection probability tending to one in the large-sample setting. [sent-11, score-0.943]
8 If a change actually occurs, the test statistic also yields an estimator of the change-point location. [sent-12, score-0.524]
9 Promising experimental results in temporal segmentation of mental tasks from BCI data and pop song indexation are presented. [sent-13, score-0.703]
10 1 Introduction The need to partition a sequence of observations into several homogeneous segments arises in many applications, ranging from speaker segmentation to pop song indexation. [sent-14, score-0.848]
11 So far, such tasks were most often dealt with using probabilistic sequence models, such as hidden Markov models [1], or their discriminative counterparts such as conditional random fields [2]. [sent-15, score-0.23]
12 These probabilistic models require a sound knowledge of the transition structure between the segments and demand careful training beforehand to yield competitive performance; when data are acquired online, inference in such models is also not straightforward (see, e. [sent-16, score-0.355]
13 Such models essentially perform multiple change-point estimation, while one is often also interested in meaningful quantitative measures for the detection of a change-point within a sample. [sent-20, score-0.221]
14 When a parametric model is available to model the distributions before and after the change, a comprehensive literature for change-point analysis has been developed, which provides optimal criteria from the maximum likelihood framework, as described in [4]. [sent-21, score-0.131]
15 Nonparametric procedures were also proposed, as reviewed in [5], but were limited to univariate data and simple settings. [sent-22, score-0.149]
16 Online counterparts have also been proposed and mostly built upon the cumulative sum scheme (see [6] for extensive references). [sent-23, score-0.115]
17 However, so far, even extensions to the case where the distribution before the change is known, and the distribution after the change is not known, remains an open problem [7]. [sent-24, score-0.286]
18 This brings to light the need to develop statistically grounded change-point analysis algorithms, working on multivariate, high-dimensional, and also structured data. [sent-25, score-0.15]
19 1 We propose here a regularized kernel-based test statistic, which allows to simultaneously provide quantitative answers to both questions: 1) is there a change-point within the sample? [sent-26, score-0.227]
20 We prove that our test statistic for change-point analysis has a false-alarm probability tending to α and a detection probability tending to one as the number of observations tends to infinity. [sent-28, score-1.218]
21 Moreover, the test statistic directly provides an accurate estimate of the change-point instant. [sent-29, score-0.347]
22 Our method readily extends to multiple change-point settings, by performing a sequence of change-point analysis in sliding windows running along the signal. [sent-30, score-0.196]
23 Usually, physical considerations allow to set the window-length to a sufficiently small length for being guaranteed that at most one change-point occurs within each window, and sufficiently large length for our decision rule to be statistically significant (typically n > 50). [sent-31, score-0.36]
24 In Section 2, we set up the framework of change-point analysis, and in Section 3, we describe how to devise a regularized kernel-based approach to the change-point problem. [sent-32, score-0.092]
25 Then, in Section 4 and in Section 5, we respectively derive the limiting distribution of our test statistic under the null hypothesis H0 : ”no change occurs“, and establish the consistency in power under the alternative HA : ”a change occurs“. [sent-33, score-1.067]
26 These theoretical results allow to build a test statistic which has provably a false-alarm probability tending to a prescribed level α, and a detection probability tending to one, as the number of observations tends to infinity. [sent-34, score-1.375]
27 Finally, in Section 7, we display the performance of our algorithm for respectively, segmentation into mental tasks from BCI data and temporal segmentation of pop songs. [sent-35, score-0.694]
28 2 Change-point analysis In this section, we outline the change-point problem, and describe formally a strategy for building change-point analysis test statistics. [sent-36, score-0.308]
29 While sharing many similarities with usual machine learning problems, the change-point problem is different. [sent-50, score-0.069]
30 Statistical hypothesis testing An important aspect of the above formulation of the changepoint problem is its natural embedding in a statistical hypothesis testing framework. [sent-51, score-0.72]
31 Let us recall briefly the main concepts in statistical hypothesis testing, in order to rephrase them within the change-point problem framework (see, e. [sent-52, score-0.262]
32 The goal is to build a decision rule to answer question 1) in the change-point problem stated above. [sent-55, score-0.134]
33 Set a false-alarm probability α with 0 < α < 1 (also called level or Type I error), whose purpose is to theoretically guarantee that P(decide HA , when H0 is true) is close to α. [sent-56, score-0.042]
34 Now, if there actually is a changepoint within the sample, one would like not to miss it, that is that the detection probability π = P(decide HA , when HA is true)—also called power and equal to one minus the Type II error—should be close to one. [sent-57, score-0.492]
35 The purpose of Sections 4-5 is to give theoretical guarantees to those practical requirements in the large-sample setting, that is when the number of observations n tends to infinity. [sent-58, score-0.236]
36 Running maximum partition strategy An efficient strategy for building change-point analysis procedures is to select the partition of the sample which yields a maximum heterogeneity between the two segments: given a sample {X1 , . [sent-59, score-0.747]
37 , Xn } and a candidate change point k with 1 < k < n, assume we may compute a measure of heterogeneity ∆n,k between the segments {X1 , . [sent-62, score-0.383]
wordName wordTfidf (topN-words)
[('ha', 0.307), ('statistic', 0.3), ('tending', 0.269), ('pop', 0.227), ('pxk', 0.227), ('occurs', 0.188), ('rue', 0.161), ('hypothesis', 0.153), ('barrault', 0.152), ('changepoint', 0.152), ('pxn', 0.152), ('change', 0.143), ('segments', 0.133), ('segmentation', 0.131), ('paris', 0.122), ('cedex', 0.121), ('detection', 0.121), ('ltci', 0.114), ('telecom', 0.114), ('testing', 0.114), ('heterogeneity', 0.107), ('prescribed', 0.107), ('bci', 0.107), ('france', 0.106), ('xn', 0.104), ('cnrs', 0.102), ('decide', 0.102), ('paristech', 0.098), ('mental', 0.095), ('null', 0.081), ('tends', 0.081), ('partition', 0.08), ('observations', 0.078), ('counterparts', 0.075), ('song', 0.074), ('strategy', 0.07), ('dealt', 0.066), ('indexation', 0.066), ('moulines', 0.066), ('limiting', 0.066), ('build', 0.065), ('temporal', 0.061), ('miss', 0.061), ('beforehand', 0.061), ('rephrase', 0.061), ('ulm', 0.061), ('sample', 0.058), ('za', 0.057), ('unlabelled', 0.057), ('xk', 0.056), ('running', 0.054), ('willow', 0.054), ('instant', 0.054), ('reviewed', 0.054), ('statistically', 0.053), ('analysis', 0.053), ('quantitative', 0.052), ('procedures', 0.052), ('establish', 0.05), ('windows', 0.049), ('tasks', 0.049), ('building', 0.049), ('within', 0.048), ('consistency', 0.048), ('francis', 0.047), ('homogeneity', 0.047), ('switches', 0.047), ('devise', 0.047), ('test', 0.047), ('speaker', 0.046), ('regularized', 0.045), ('brings', 0.044), ('eric', 0.044), ('demand', 0.043), ('comprehensive', 0.043), ('univariate', 0.043), ('purpose', 0.042), ('sound', 0.041), ('upon', 0.04), ('sequence', 0.04), ('minus', 0.04), ('acquired', 0.04), ('homogeneous', 0.039), ('provably', 0.038), ('online', 0.037), ('careful', 0.037), ('outline', 0.036), ('fisher', 0.036), ('rule', 0.036), ('power', 0.036), ('answers', 0.035), ('sharing', 0.035), ('maximum', 0.035), ('requirements', 0.035), ('considerations', 0.035), ('similarities', 0.034), ('aspect', 0.034), ('actually', 0.034), ('consists', 0.033), ('answer', 0.033)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000001 111 nips-2008-Kernel Change-point Analysis
Author: Zaïd Harchaoui, Eric Moulines, Francis R. Bach
Abstract: We introduce a kernel-based method for change-point analysis within a sequence of temporal observations. Change-point analysis of an unlabelled sample of observations consists in, first, testing whether a change in the distribution occurs within the sample, and second, if a change occurs, estimating the change-point instant after which the distribution of the observations switches from one distribution to another different distribution. We propose a test statistic based upon the maximum kernel Fisher discriminant ratio as a measure of homogeneity between segments. We derive its limiting distribution under the null hypothesis (no change occurs), and establish the consistency under the alternative hypothesis (a change occurs). This allows to build a statistical hypothesis testing procedure for testing the presence of a change-point, with a prescribed false-alarm probability and detection probability tending to one in the large-sample setting. If a change actually occurs, the test statistic also yields an estimator of the change-point location. Promising experimental results in temporal segmentation of mental tasks from BCI data and pop song indexation are presented. 1
2 0.15332706 189 nips-2008-Rademacher Complexity Bounds for Non-I.I.D. Processes
Author: Mehryar Mohri, Afshin Rostamizadeh
Abstract: This paper presents the first Rademacher complexity-based error bounds for noni.i.d. settings, a generalization of similar existing bounds derived for the i.i.d. case. Our bounds hold in the scenario of dependent samples generated by a stationary β-mixing process, which is commonly adopted in many previous studies of noni.i.d. settings. They benefit from the crucial advantages of Rademacher complexity over other measures of the complexity of hypothesis classes. In particular, they are data-dependent and measure the complexity of a class of hypotheses based on the training sample. The empirical Rademacher complexity can be estimated from such finite samples and lead to tighter generalization bounds. We also present the first margin bounds for kernel-based classification in this non-i.i.d. setting and briefly study their convergence.
3 0.076314412 88 nips-2008-From Online to Batch Learning with Cutoff-Averaging
Author: Ofer Dekel
Abstract: We present cutoff averaging, a technique for converting any conservative online learning algorithm into a batch learning algorithm. Most online-to-batch conversion techniques work well with certain types of online learning algorithms and not with others, whereas cutoff averaging explicitly tries to adapt to the characteristics of the online algorithm being converted. An attractive property of our technique is that it preserves the efficiency of the original online algorithm, making it appropriate for large-scale learning problems. We provide a statistical analysis of our technique and back our theoretical claims with experimental results. 1
4 0.073610514 138 nips-2008-Modeling human function learning with Gaussian processes
Author: Thomas L. Griffiths, Chris Lucas, Joseph Williams, Michael L. Kalish
Abstract: Accounts of how people learn functional relationships between continuous variables have tended to focus on two possibilities: that people are estimating explicit functions, or that they are performing associative learning supported by similarity. We provide a rational analysis of function learning, drawing on work on regression in machine learning and statistics. Using the equivalence of Bayesian linear regression and Gaussian processes, we show that learning explicit rules and using similarity can be seen as two views of one solution to this problem. We use this insight to define a Gaussian process model of human function learning that combines the strengths of both approaches. 1
5 0.070916899 147 nips-2008-Multiscale Random Fields with Application to Contour Grouping
Author: Longin J. Latecki, Chengen Lu, Marc Sobel, Xiang Bai
Abstract: We introduce a new interpretation of multiscale random fields (MSRFs) that admits efficient optimization in the framework of regular (single level) random fields (RFs). It is based on a new operator, called append, that combines sets of random variables (RVs) to single RVs. We assume that a MSRF can be decomposed into disjoint trees that link RVs at different pyramid levels. The append operator is then applied to map RVs in each tree structure to a single RV. We demonstrate the usefulness of the proposed approach on a challenging task involving grouping contours of target shapes in images. It provides a natural representation of multiscale contour models, which is needed in order to cope with unstable contour decompositions. The append operator allows us to find optimal image segment labels using the classical framework of relaxation labeling. Alternative methods like Markov Chain Monte Carlo (MCMC) could also be used.
6 0.0707912 191 nips-2008-Recursive Segmentation and Recognition Templates for 2D Parsing
7 0.070510797 47 nips-2008-Clustered Multi-Task Learning: A Convex Formulation
8 0.070334114 42 nips-2008-Cascaded Classification Models: Combining Models for Holistic Scene Understanding
9 0.06657812 180 nips-2008-Playing Pinball with non-invasive BCI
10 0.065700755 208 nips-2008-Shared Segmentation of Natural Scenes Using Dependent Pitman-Yor Processes
11 0.054827314 65 nips-2008-Domain Adaptation with Multiple Sources
12 0.052853823 21 nips-2008-An Homotopy Algorithm for the Lasso with Online Observations
13 0.047658 72 nips-2008-Empirical performance maximization for linear rank statistics
14 0.046400566 97 nips-2008-Hierarchical Fisher Kernels for Longitudinal Data
15 0.043027002 112 nips-2008-Kernel Measures of Independence for non-iid Data
16 0.042778023 76 nips-2008-Estimation of Information Theoretic Measures for Continuous Random Variables
17 0.04260198 148 nips-2008-Natural Image Denoising with Convolutional Networks
18 0.041648269 61 nips-2008-Diffeomorphic Dimensionality Reduction
19 0.039825097 159 nips-2008-On Bootstrapping the ROC Curve
20 0.037280079 203 nips-2008-Scalable Algorithms for String Kernels with Inexact Matching
topicId topicWeight
[(0, -0.14), (1, -0.007), (2, -0.004), (3, -0.012), (4, -0.017), (5, 0.003), (6, -0.009), (7, -0.033), (8, 0.077), (9, 0.029), (10, 0.006), (11, 0.016), (12, 0.078), (13, -0.051), (14, -0.008), (15, 0.082), (16, 0.03), (17, -0.015), (18, -0.034), (19, 0.036), (20, 0.011), (21, 0.01), (22, 0.129), (23, 0.07), (24, 0.033), (25, 0.058), (26, 0.021), (27, 0.095), (28, -0.047), (29, -0.01), (30, 0.058), (31, -0.006), (32, -0.066), (33, 0.073), (34, -0.061), (35, 0.017), (36, 0.1), (37, -0.003), (38, 0.049), (39, -0.15), (40, -0.119), (41, 0.08), (42, 0.129), (43, 0.078), (44, -0.072), (45, 0.087), (46, -0.022), (47, 0.082), (48, 0.033), (49, 0.106)]
simIndex simValue paperId paperTitle
same-paper 1 0.94178748 111 nips-2008-Kernel Change-point Analysis
Author: Zaïd Harchaoui, Eric Moulines, Francis R. Bach
Abstract: We introduce a kernel-based method for change-point analysis within a sequence of temporal observations. Change-point analysis of an unlabelled sample of observations consists in, first, testing whether a change in the distribution occurs within the sample, and second, if a change occurs, estimating the change-point instant after which the distribution of the observations switches from one distribution to another different distribution. We propose a test statistic based upon the maximum kernel Fisher discriminant ratio as a measure of homogeneity between segments. We derive its limiting distribution under the null hypothesis (no change occurs), and establish the consistency under the alternative hypothesis (a change occurs). This allows to build a statistical hypothesis testing procedure for testing the presence of a change-point, with a prescribed false-alarm probability and detection probability tending to one in the large-sample setting. If a change actually occurs, the test statistic also yields an estimator of the change-point location. Promising experimental results in temporal segmentation of mental tasks from BCI data and pop song indexation are presented. 1
2 0.59145665 189 nips-2008-Rademacher Complexity Bounds for Non-I.I.D. Processes
Author: Mehryar Mohri, Afshin Rostamizadeh
Abstract: This paper presents the first Rademacher complexity-based error bounds for noni.i.d. settings, a generalization of similar existing bounds derived for the i.i.d. case. Our bounds hold in the scenario of dependent samples generated by a stationary β-mixing process, which is commonly adopted in many previous studies of noni.i.d. settings. They benefit from the crucial advantages of Rademacher complexity over other measures of the complexity of hypothesis classes. In particular, they are data-dependent and measure the complexity of a class of hypotheses based on the training sample. The empirical Rademacher complexity can be estimated from such finite samples and lead to tighter generalization bounds. We also present the first margin bounds for kernel-based classification in this non-i.i.d. setting and briefly study their convergence.
3 0.51552022 169 nips-2008-Online Models for Content Optimization
Author: Deepak Agarwal, Bee-chung Chen, Pradheep Elango, Nitin Motgi, Seung-taek Park, Raghu Ramakrishnan, Scott Roy, Joe Zachariah
Abstract: We describe a new content publishing system that selects articles to serve to a user, choosing from an editorially programmed pool that is frequently refreshed. It is now deployed on a major Yahoo! portal, and selects articles to serve to hundreds of millions of user visits per day, significantly increasing the number of user clicks over the original manual approach, in which editors periodically selected articles to display. Some of the challenges we face include a dynamic content pool, short article lifetimes, non-stationary click-through rates, and extremely high traffic volumes. The fundamental problem we must solve is to quickly identify which items are popular (perhaps within different user segments), and to exploit them while they remain current. We must also explore the underlying pool constantly to identify promising alternatives, quickly discarding poor performers. Our approach is based on tracking per article performance in near real time through online models. We describe the characteristics and constraints of our application setting, discuss our design choices, and show the importance and effectiveness of coupling online models with a randomization procedure. We discuss the challenges encountered in a production online content-publishing environment and highlight issues that deserve careful attention. Our analysis of this application also suggests a number of future research avenues. 1
4 0.46151334 41 nips-2008-Breaking Audio CAPTCHAs
Author: Jennifer Tam, Jiri Simsa, Sean Hyde, Luis V. Ahn
Abstract: CAP TCHAs are computer-generated tests that humans can pass but current computer systems cannot. CAP TCHAs provide a method for automatically distinguishing a human from a computer program, and therefore can protect Web services from abuse by so-called “bots.” Most CAP TCHAs consist of distorted images, usually text, for which a user must provide some description. Unfortunately, visual CAP TCHAs limit access to the millions of visually impaired people using the Web. Audio CAP TCHAs were created to solve this accessibility issue; however, the security of audio CAP TCHAs was never formally tested. Some visual CAP TCHAs have been broken using machine learning techniques, and we propose using similar ideas to test the security of audio CAP TCHAs. Audio CAP TCHAs are generally composed of a set of words to be identified, layered on top of noise. We analyzed the security of current audio CAP TCHAs from popular Web sites by using AdaBoost, SVM, and k-NN, and achieved correct solutions for test samples with accuracy up to 71%. Such accuracy is enough to consider these CAPTCHAs broken. Training several different machine learning algorithms on different types of audio CAP TCHAs allowed us to analyze the strengths and weaknesses of the algorithms so that we could suggest a design for a more robust audio CAPTCHA. 1 Int rod uct i o n CAP TCHAs [1] are automated tests designed to tell computers and humans apart by presenting users with a problem that humans can solve but current computer programs cannot. Because CAPTCHAs can distinguish between humans and computers with high probability, they are used for many different security applications: they prevent bots from voting continuously in online polls, automatically registering for millions of spam email accounts, automatically purchasing tickets to buy out an event, etc. Once a CAP TCHA is broken (i.e., computer programs can successfully pass the test), bots can impersonate humans and gain access to services that they should not. Therefore, it is important for CAP TCHAs to be secure. To pass the typical visual CAP TCHA, a user must correctly type the characters displayed in an image of distorted text. Many visual CAP TCHAs have been broken with machine learning techniques [2]-[3], though some remain secure against such attacks. Because visually impaired users who surf the Web using screen-reading programs cannot see this type of CAPTCHA, audio CAP TCHAs were created. Typical audio CAP TCHAs consist of one or several speakers saying letters or digits at randomly spaced intervals. A user must correctly identify the digits or characters spoken in the audio file to pass the CAP TCHA. To make this test difficult for current computer systems, specifically automatic speech recognition (ASR) programs, background noise is injected into the audio files. Since no official evaluation of existing audio CAP TCHAs has been reported, we tested the security of audio CAP TCHAs used by many popular Web sites by running machine learning experiments designed to break them. In the next section, we provide an overview of the literature related to our project. Section 3 describes our methods for creating training data, and section 4 describes how we create classifiers that can recognize letters, digits, and noise. In section 5, we discuss how we evaluated our methods on widely used audio CAP TCHAs and we give our results. In particular, we show that the audio CAP TCHAs used by sites such as Google and Digg are susceptible to machine learning attacks. Section 6 mentions the proposed design of a new more secure audio CAP TCHA based on our findings. 2 Lit erat u r e r ev i ew To break the audio CAP TCHAs, we derive features from the CAP TCHA audio and use several machine learning techniques to perform ASR on segments of the CAPTCHA. There are many popular techniques for extracting features from speech. The three techniques we use are mel-frequency cepstral coefficients (MFCC), perceptual linear prediction (PLP), and relative spectral transform-PLP (RAS TA-PLP). MFCC is one of the most popular speech feature representations used. Similar to a fast Fourier transform (FF T), MFCC transforms an audio file into frequency bands, but (unlike FF T) MFCC uses mel-frequency bands, which are better for approximating the range of frequencies humans hear. PLP was designed to extract speaker-independent features from speech [4]. Therefore, by using PLP and a variant such as RAS TA-PL P, we were able to train our classifiers to recognize letters and digits independently of who spoke them. Since many different people recorded the digits used in one of the types of audio CAP TCHAs we tested, PLP and RAS TA-PLP were needed to extract the features that were most useful for solving them. In [4]-[5], the authors conducted experiments on recognizing isolated digits in the presence of noise using both PLP and RAS TA-PL P. However, the noise used consisted of telephone or microphone static caused by recording in different locations. The audio CAP TCHAs we use contain this type of noise, as well as added vocal noise and/or music, which is supposed to make the automated recognition process much harder. The authors of [3] emphasize how many visual CAP TCHAs can be broken by successfully splitting the task into two smaller tasks: segmentation and recognition. We follow a similar approach in that we first automatically split the audio into segments, and then we classify these segments as noise or words. In early March 2008, concurrent to our work, the blog of Wintercore Labs [6] claimed to have successfully broken the Google audio CAP TCHA. After reading their Web article and viewing the video of how they solve the CAP TCHAs, we are unconvinced that the process is entirely automatic, and it is unclear what their exact pass rate is. Because we are unable to find any formal technical analysis of this program, we can neither be sure of its accuracy nor the extent of its automation. 3 Cr e at i o n of tra i n i n g dat a Since automated programs can attempt to pass a CAPTCHA repeatedly, a CAPTCHA is essentially broken when a program can pass it more than a non-trivial fraction of the time; e.g., a 5% pass rate is enough. Our approach to breaking the audio CAP TCHAs began by first splitting the audio files into segments of noise or words: for our experiments, the words were spoken letters or digits. We used manual transcriptions of the audio CAP TCHAs to get information regarding the location of each spoken word within the audio file. We were able to label our segments accurately by using this information. We gathered 1,000 audio CAP TCHAs from each of the following Web sites: google.com, digg.com, and an older version of the audio CAP TCHA in recaptcha.net. Each of the CAP TCHAs was annotated with the information regarding letter/digit locations provided by the manual transcriptions. For each type of CAPTCHA, we randomly selected 900 samples for training and used the remaining 100 for testing. Using the digit/letter location information provided in the manual CAP TCHA transcriptions, each training CAP TCHA is divided into segments of noise, the letters a-z, or the digits 0-9, and labeled as such. We ignore the annotation information of the CAP TCHAs we use for testing, and therefore we cannot identify the size of those segments. Instead, each test CAP TCHA is divided into a number of fixed-size segments. The segments with the highest energy peaks are then classified using machine learning techniques (Figure 1). Since the size of a feature vector extracted from a segment generally depends on the size of the segment, using fixed-size segments allows each segment to be described with a feature vector of the same length. We chose the window size by listening to a few training segments and adjusted accordingly to ensure that the segment contained the entire digit/letter. There is undoubtedly a more optimal way of selecting the window size, however, we were still able to break the three CAP TCHAs we tested with our method. Figure 1: A test audio CAP TCHA with the fixed-size segments containing the highest energy peaks highlighted. The information provided in the manual transcriptions of the audio CAP TCHAs contains a list of the time intervals within which words are spoken. However, these intervals are of variable size and the word might be spoken anywhere within this interval. To provide fixedsize segments for training, we developed the following heuristic. First, divide each file into variable-size segments using the time intervals provided and label each segment accordingly. Then, within each segment, detect the highest energy peak and return its fixed-size neighborhood labeled with the current segment’s label. This heuristic achieved nearly perfect labeling accuracy for the training set. Rare mistakes occurred when the highest energy peak of a digit or letter segment corresponded to noise rather than to a digit or letter. To summarize this subsection, an audio file is transformed into a set of fixed-size segments labeled as noise, a digit between 0 and 9, or a letter between a and z. These segments are then used for training. Classifiers are trained for one type of CAPTCHA at a time. 4 C l a s s i f i e r con s t ru ct i o n From the training data we extracted five sets of features using twelve MFCCs and twelfth- order spectral (SPEC) and cepstral (CEPS) coefficients from PLP Matlab functions for extracting these features were provided online Voicebox package. We use AdaBoost, SVM, and k-NN algorithms digit and letter recognition. We detail our implementation of following subsections. 4 .1 and RAS TA-PL P. The at [7] and as part of the to implement automated each algorithm in the AdaBoost Using decision stumps as weak classifiers for AdaBoost, anywhere from 11 to 37 ensemble classifiers are built. The number of classifiers built depends on which type of CAPTCHA we are solving. Each classifier trains on all the segments associated with that type of CAP TCHA, and for the purpose of building a single classifier, segments are labeled by either -1 (negative example) or +1 (positive example). Using cross-validation, we choose to use 50 iterations for our AdaBoost algorithm. A segment can then be classified as a particular letter, digit, or noise according to the ensemble classifier that outputs the number closest to 1. 4 .2 S u p p o rt v e ct o r m a c h i n e To conduct digit recognition with SVM, we used the C++ implementations of libSVM [8] version 2.85 with C-SMV and RBF kernel. First, all feature values are scaled to the range of -1 to 1 as suggested by [8]. The scale parameters are stored so that test samples can be scaled accordingly. Then, a single multiclass classifier is created for each set of features using all the segments for a particular type of CAPTCHA. We use cross-validation and grid search to discover the optimal slack penalty (C=32) and kernel parameter (γ=0.011). 4 .3 k - n e a re st n e i g h b o r ( k - N N ) We use k-NN as our final method for classifying digits. For each type of CAP TCHA, five different classifiers are created by using all of the training data and the five sets of features associated with that particular type of CAP TCHA. Again we use cross-validation to discover the optimal parameter, in this case k=1. We use Euclidian distance as our distance metric. 5 Ass e s sm e n t of cu rre n t a ud i o CAPTCHAs Our method for solving CAP TCHAs iteratively extracts an audio segment from a CAP TCHA, inputs the segment to one of our digit or letter recognizers, and outputs the label for that segment. We continue this process until the maximum solution size is reached or there are no unlabeled segments left. Some of the CAPTCHAs we evaluated have solutions that vary in length. Our method ensures that we get solutions of varying length that are never longer than the maximum solution length. A segment to be classified is identified by taking the neighborhood of the highest energy peak of an as yet unlabeled part of the CAP TCHA. Once a prediction of the solution to the CAPTCHA is computed, it is compared to the true solution. Given that at least one of the audio CAP TCHAs allows users to make a mistake in one of the digits (e.g., reCAPTCHA), we compute the pass rate for each of the different types of CAPTCHAs with all of the following conditions: • The prediction matches the true solution exactly. • Inserting one digit into the prediction would make it match the solution exactly. • Replacing one digit in the prediction would make it match the solution exactly. • Removing one digit from the prediction would make it match the solution exactly. However, since we are only sure that these conditions apply to reCAPTCHA audio CAP TCHAs, we also calculate the percentage of exact solution matches in our results for each type of audio CAP TCHA. These results are described in the following subsections. 5 .1 Goog le Google audio CAP TCHAs consist of one speaker saying random digits 0-9, the phrase “once again,” followed by the exact same recorded sequence of digits originally presented. The background noise consists of human voices speaking backwards at varying volumes. A solution can range in length from five to eight words. We set our classifier to find the 12 loudest segments and classify these segments as digits or noise. Because the phrase “once again” marks the halfway point of the CAPTCHA, we preprocessed the audio to only serve this half of the CAP TCHA to our classifiers. It is important to note, however, that the classifiers were always able to identify the segment containing “once again,” and these segments were identified before all other segments. Therefore, if necessary, we could have had our system cut the file in half after first labeling this segment. For AdaBoost, we create 12 classifiers: one classifier for each digit, one for noise, and one for the phrase “once again.” Our results ( Table 1) show that at best we achieved a 90% pass rate using the “one mistake” passing conditions and a 66% exact solution match rate. Using SVM and the “one mistake” passing conditions, at best we achieve a 92% pass rate and a 67% exact solution match. For k-NN, the “one mistake” pass rate is 62% and the exact solution match rate is 26%. Table 1: Google audio CAP TCHA results: Maximum 67% accuracy was achieved by SVM. Classifiers Used AdaBoost SVM k-NN one mistake exact match one mistake exact match 88% 61% 92% 67% 30% 1% PLPSPEC 90% 66% 90% 67% 60% 26% PLPCEPS 90% 66% 92% 67% 62% 23% RAS TAPLPSPEC 88% 48% 90% 61% 29% 1% RAS TAPLPCEPS 5 .2 exact match MFCC Features Used One mistake 90% 63% 92% 67% 33% 2% Digg Digg CAP TCHAs also consist of one speaker, in this case saying a random combination of letters and digits. The background noise consists of static or what sounds like trickling water and is not continuous throughout the entire file. We noticed in our training data that the following characters were never present in a solution: 0, 1, 2, 5, 7, 9, i, o, z. Since the Digg audio CAPTCHA is also the verbal transcription of the visual CAP TCHA, we believe that these characters are excluded to avoid confusion between digits and letters that are similar in appearance. The solution length varies between three and six words. Using AdaBoost, we create 28 classifiers: one classifier for each digit or letter that appears in our training data and one classifier for noise. Perhaps because we had fewer segments to train with and there was a far higher proportion of noise segments, AdaBoost failed to produce any correct solutions. We believe that the overwhelming number of negative training examples versus the small number of positive training samples used to create each decision stump severely affected AdaBoost’s ability to classify audio segments correctly. A histogram of the training samples is provided in Figure 2 to illustrate the amount of training data available for each character. When using SVM, the best feature set passed with 96% using “one mistake” passing conditions and passed with 71% when matching the solution exactly. For k-NN, the best feature set produced a 90% “one mistake” pass rate and a 49% exact solution match. Full results can be found in Table 2. Table 2: Digg audio CAP TCHA results: Maximum 71% accuracy was achieved by SVM. Classifiers Used AdaBoost SVM k-NN exact match one mistake exact match one mistake exact match MFCC - - 96% 71% 89% 49% PLPSPEC - - 94% 65% 90% 47% PLPCEPS - - 96% 71% 64% 17% RAS TAPLPSPEC - - 17% 3% 67% 17% RAS TAPLPCEPS Features Used one mistake - - 96% 71% 82% 34% 1000 900 800 # of Segments 700 600 500 400 300 200 100 y x v w t u r s q p n m l j k h f g e d c b a 8 6 4 3 noise 0 Segment Label Figure 2: Digg CAP TCHA training data distribution. 5 .3 reC A P T C H A The older version of reCAPTCHA’s audio CAP TCHAs we tested consist of several speakers who speak random digits. The background noise consists of human voices speaking backwards at varying volumes. The solution is always eight digits long. For AdaBoost, we create 11 classifiers: one classifier for each digit and one classifier for noise. Because we know that the reCAPTCHA passing conditions are the “one mistake” passing conditions, SVM produces our best pass rate of 58%. Our best exact match rate is 45% ( Table 3). Table 3: reCAPTCHA audio CAP TCHA results: Maximum 45% accuracy was achieved by SVM. Classifiers Used AdaBoost SVM k-NN one mistake exact match one mistake exact match 18% 6% 56% 43% 22% 11% PLPSPEC 27% 10% 58% 39% 43% 25% PLPCEPS 23% 10% 56% 45% 29% 14% RAS TAPLPSPEC 9% 3% 36% 18% 24% 4% RAS TAPLPCEPS 6 exact match MFCC Features Used one mistake 9% 3% 46% 30% 32% 12% Prop ert i e s of w ea k ver s u s st ro n g CAPT CHAs From our results, we note that the easiest CAP TCHAs to break were from Digg. Google had the next strongest CAP TCHAs followed by the strongest from reCAPTCHA. Although the Digg CAP TCHAs have the largest vocabulary, giving us less training data per label, the same woman recorded them all. More importantly, the same type of noise is used throughout the entire CAPTCHA. The noise sounds like running water and static which sounds very different from the human voice and does not produce the same energy spikes needed to locate segments, therefore making segmentation quite easy. The CAP TCHAs from Google and reCAPTCHA used other human voices for background noise, making segmentation much more difficult. Although Google used a smaller vocabulary than Digg and also only used one speaker, Google’s background noise made the CAP TCHA more difficult to solve. After listening to a few of Google’s CAP TCHAs, we noticed that although the background noise consisted of human voices, the same background noise was repeated. reCAP TCHA had similar noise to Google, but they had a larger selection of noise thus making it harder to learn. reCAP TCHA also has the longest solution length making it more difficult to get perfectly correct. Finally, reCAPTCHA used many different speakers causing it to be the strongest CAP TCHA of the three we tested. In conclusion, an audio CAP TCHA that consists of a finite vocabulary and background noise should have multiple speakers and noise similar to the speakers. 7 Recomm e n d at i o n s f or creat i n g st ro n g e r aud i o CAPTCHAs Due to our success in solving audio CAP TCHAs, we have decided to start developing new audio CAP TCHAs that our methods, and machine learning methods in general, will be less likely to solve. From our experiments, we note that CAP TCHAs containing longer solutions and multiple speakers tend to be more difficult to solve. Also, because our methods depend on the amount of training data we have, having a large vocabulary would make it more difficult to collect enough training data. Already since obtaining these results, reCAPTCHA.net has updated their audio CAP TCHA to contain more distortions and a larger vocabulary: the digits 0 through 99. In designing a new audio CAP TCHA we are also concerned with the human pass rate. The current human pass rate for the reCAPTCHA audio CAP TCHAs is only 70%. To develop an audio CAP TCHA with an improved human pass rate, we plan to take advantage of the human mind’s ability to understand distorted audio through context clues. By listening to a phrase instead of to random isolated words, humans are better able to decipher distorted utterances because they are familiar with the phrase or can use contextual clues to decipher the distorted audio. Using this idea, the audio for our new audio CAP TCHA will be taken from old-time radio programs in which the poor quality of the audio makes transcription by ASR systems difficult. Users will be presented with an audio clip consisting of a 4-6 word phrase. Half of the CAPTCHA consists of words, which validate a user to be human, while the other half of the words need to be transcribed. This is the same idea behind the visual reCAP TCHA that is currently digitizing text on which OCR fails. We expect that this new audio CAP TCHA will be more secure than the current version and easier for humans to pass. Initial experiments using this idea show this to be true [9]. 8 Co n c l u s i o n We have succeeded in “breaking” three different types of widely used audio CAP TCHAs, even though these were developed with the purpose of defeating attacks by machine learning techniques. We believe our results can be improved by selecting optimal segment sizes, but that is unnecessary given our already high success rate. For our experiments, segment sizes were not chosen in a special way; occasionally yielding results in which a segment only contained half of a word, causing our prediction to contain that particular word twice. We also believe that the AdaBoost results can be improved, particularly for the Digg audio CAP TCHAs, by ensuring that the number of negative training samples is closer to the number of positive training samples. We have shown that our approach is successful and can be used with many different audio CAP TCHAs that contain small finite vocabularies. A ck n o w l e d g m e n t s This work was partially supported by generous gifts from the Heinz Endowment, by an equipment grant from Intel Corporation, and by the Army Research Office through grant number DAAD19-02-1-0389 to CyLab at Carnegie Mellon University. Luis von Ahn was partially supported by a Microsoft Research New Faculty Fellowship and a MacArthur Fellowship. Jennifer Tam was partially supported by a Google Anita Borg Scholarship. R e f e re n c e s [1] L. von Ahn, M. Blum, and J. Langford. “Telling Humans and Computers Apart Automatically,” Communication s of the ACM, vol. 47, no. 2, pp. 57-60, Feb. 2004. [2] G. Mori and J. Malik. “Recognizing Objects in Adversarial Clutter: Breaking a Visual CAPTCHA,” In Computer Vision and Pattern Recognition CVPR'03, June 2003. [3] K. Chellapilla, and P. Simard, “ U sing Machine Learning to Break Visual Human Interactio n Proofs (HIP s),” Advances in Neural Information P rocessing Systems 17, Neural Info rmatio n P rocessing Systems (NIPS'2004), MIT Press. [4] H. Hermansk y, “ Perceptual Linear Predictive (PL P) Analysis of Speech,” J. Acoust. Soc. Am., vol. 87, no. 4, pp. 1738-1752, Apr. 1990. [5] H. Hermansk y, N. Morgan, A. Bayya, and P. Kohn. “RASTA-PL P Speech Analysi s Technique,” In P roc. IEEE Int’l Conf. Acoustics, Speech & Signal Processing, vol. 1, pp. 121124, San Francisco, 1992. [6] R. Santamarta. “Breaking Gmail ’s Audio Captcha,” http://blog.wintercore.com/?p=11, 2008. [7] D. Ell is. “ P L P and RASTA (and MFCC, and inversion) in Matlab using melfcc.m and invmelfcc.m,” http:/ /ww w.ee.columbia.edu/~dpwe/resources/matlab/rastamat/, 2006. [8] C. Chang and C. Lin. LIBSVM: a library for support vector machines, 2001. Software available at http: //ww w.csie.ntu.edu.tw/~cjlin/libsvm [9] A. Schlaikjer. “ A Dual-Use Speech CA PTCHA: Aiding Visually Impaired Web Users while Providing Transcriptions of Audio Streams,” Technical Report CMU-LTI-07-014, Carnegie Mellon Universi t y. November 2007.
5 0.44638714 208 nips-2008-Shared Segmentation of Natural Scenes Using Dependent Pitman-Yor Processes
Author: Erik B. Sudderth, Michael I. Jordan
Abstract: We develop a statistical framework for the simultaneous, unsupervised segmentation and discovery of visual object categories from image databases. Examining a large set of manually segmented scenes, we show that object frequencies and segment sizes both follow power law distributions, which are well modeled by the Pitman–Yor (PY) process. This nonparametric prior distribution leads to learning algorithms which discover an unknown set of objects, and segmentation methods which automatically adapt their resolution to each image. Generalizing previous applications of PY processes, we use Gaussian processes to discover spatially contiguous segments which respect image boundaries. Using a novel family of variational approximations, our approach produces segmentations which compare favorably to state-of-the-art methods, while simultaneously discovering categories shared among natural scenes. 1
6 0.44584018 211 nips-2008-Simple Local Models for Complex Dynamical Systems
7 0.41797194 236 nips-2008-The Mondrian Process
8 0.41711372 138 nips-2008-Modeling human function learning with Gaussian processes
9 0.37977964 161 nips-2008-On the Complexity of Linear Prediction: Risk Bounds, Margin Bounds, and Regularization
10 0.3722339 191 nips-2008-Recursive Segmentation and Recognition Templates for 2D Parsing
11 0.35197341 147 nips-2008-Multiscale Random Fields with Application to Contour Grouping
12 0.34865895 110 nips-2008-Kernel-ARMA for Hand Tracking and Brain-Machine interfacing During 3D Motor Control
13 0.34643608 221 nips-2008-Stochastic Relational Models for Large-scale Dyadic Data using MCMC
14 0.34018552 88 nips-2008-From Online to Batch Learning with Cutoff-Averaging
15 0.33970419 42 nips-2008-Cascaded Classification Models: Combining Models for Holistic Scene Understanding
16 0.33791372 154 nips-2008-Nonparametric Bayesian Learning of Switching Linear Dynamical Systems
17 0.33441922 185 nips-2008-Privacy-preserving logistic regression
18 0.32837719 100 nips-2008-How memory biases affect information transmission: A rational analysis of serial reproduction
19 0.32797509 148 nips-2008-Natural Image Denoising with Convolutional Networks
20 0.32678053 180 nips-2008-Playing Pinball with non-invasive BCI
topicId topicWeight
[(6, 0.061), (7, 0.054), (12, 0.032), (28, 0.132), (57, 0.048), (63, 0.019), (77, 0.031), (78, 0.497), (83, 0.039)]
simIndex simValue paperId paperTitle
1 0.82120699 52 nips-2008-Correlated Bigram LSA for Unsupervised Language Model Adaptation
Author: Yik-cheung Tam, Tanja Schultz
Abstract: We present a correlated bigram LSA approach for unsupervised LM adaptation for automatic speech recognition. The model is trained using efficient variational EM and smoothed using the proposed fractional Kneser-Ney smoothing which handles fractional counts. We address the scalability issue to large training corpora via bootstrapping of bigram LSA from unigram LSA. For LM adaptation, unigram and bigram LSA are integrated into the background N-gram LM via marginal adaptation and linear interpolation respectively. Experimental results on the Mandarin RT04 test set show that applying unigram and bigram LSA together yields 6%–8% relative perplexity reduction and 2.5% relative character error rate reduction which is statistically significant compared to applying only unigram LSA. On the large-scale evaluation on Arabic, 3% relative word error rate reduction is achieved which is also statistically significant. 1
same-paper 2 0.80079478 111 nips-2008-Kernel Change-point Analysis
Author: Zaïd Harchaoui, Eric Moulines, Francis R. Bach
Abstract: We introduce a kernel-based method for change-point analysis within a sequence of temporal observations. Change-point analysis of an unlabelled sample of observations consists in, first, testing whether a change in the distribution occurs within the sample, and second, if a change occurs, estimating the change-point instant after which the distribution of the observations switches from one distribution to another different distribution. We propose a test statistic based upon the maximum kernel Fisher discriminant ratio as a measure of homogeneity between segments. We derive its limiting distribution under the null hypothesis (no change occurs), and establish the consistency under the alternative hypothesis (a change occurs). This allows to build a statistical hypothesis testing procedure for testing the presence of a change-point, with a prescribed false-alarm probability and detection probability tending to one in the large-sample setting. If a change actually occurs, the test statistic also yields an estimator of the change-point location. Promising experimental results in temporal segmentation of mental tasks from BCI data and pop song indexation are presented. 1
3 0.78262633 124 nips-2008-Load and Attentional Bayes
Author: Peter Dayan
Abstract: Selective attention is a most intensively studied psychological phenomenon, rife with theoretical suggestions and schisms. A critical idea is that of limited capacity, the allocation of which has produced continual conflict about such phenomena as early and late selection. An influential resolution of this debate is based on the notion of perceptual load (Lavie, 2005), which suggests that low-load, easy tasks, because they underuse the total capacity of attention, mandatorily lead to the processing of stimuli that are irrelevant to the current attentional set; whereas high-load, difficult tasks grab all resources for themselves, leaving distractors high and dry. We argue that this theory presents a challenge to Bayesian theories of attention, and suggest an alternative, statistical, account of key supporting data. 1
4 0.71544594 82 nips-2008-Fast Computation of Posterior Mode in Multi-Level Hierarchical Models
Author: Liang Zhang, Deepak Agarwal
Abstract: Multi-level hierarchical models provide an attractive framework for incorporating correlations induced in a response variable that is organized hierarchically. Model fitting is challenging, especially for a hierarchy with a large number of nodes. We provide a novel algorithm based on a multi-scale Kalman filter that is both scalable and easy to implement. For Gaussian response, we show our method provides the maximum a-posteriori (MAP) parameter estimates; for non-Gaussian response, parameter estimation is performed through a Laplace approximation. However, the Laplace approximation provides biased parameter estimates that is corrected through a parametric bootstrap procedure. We illustrate through simulation studies and analyses of real world data sets in health care and online advertising.
5 0.40840271 216 nips-2008-Sparse probabilistic projections
Author: Cédric Archambeau, Francis R. Bach
Abstract: We present a generative model for performing sparse probabilistic projections, which includes sparse principal component analysis and sparse canonical correlation analysis as special cases. Sparsity is enforced by means of automatic relevance determination or by imposing appropriate prior distributions, such as generalised hyperbolic distributions. We derive a variational Expectation-Maximisation algorithm for the estimation of the hyperparameters and show that our novel probabilistic approach compares favourably to existing techniques. We illustrate how the proposed method can be applied in the context of cryptoanalysis as a preprocessing tool for the construction of template attacks. 1
6 0.39537686 147 nips-2008-Multiscale Random Fields with Application to Contour Grouping
7 0.39085129 36 nips-2008-Beyond Novelty Detection: Incongruent Events, when General and Specific Classifiers Disagree
8 0.38770419 98 nips-2008-Hierarchical Semi-Markov Conditional Random Fields for Recursive Sequential Data
9 0.38768265 20 nips-2008-An Extended Level Method for Efficient Multiple Kernel Learning
10 0.38417217 169 nips-2008-Online Models for Content Optimization
11 0.36967555 15 nips-2008-Adaptive Martingale Boosting
12 0.36495873 142 nips-2008-Multi-Level Active Prediction of Useful Image Annotations for Recognition
13 0.36269742 66 nips-2008-Dynamic visual attention: searching for coding length increments
14 0.36165833 4 nips-2008-A Scalable Hierarchical Distributed Language Model
15 0.35987914 205 nips-2008-Semi-supervised Learning with Weakly-Related Unlabeled Data : Towards Better Text Categorization
16 0.3580423 130 nips-2008-MCBoost: Multiple Classifier Boosting for Perceptual Co-clustering of Images and Visual Features
17 0.35588694 10 nips-2008-A rational model of preference learning and choice prediction by children
18 0.35540187 79 nips-2008-Exploring Large Feature Spaces with Hierarchical Multiple Kernel Learning
19 0.35383341 17 nips-2008-Algorithms for Infinitely Many-Armed Bandits
20 0.35338798 167 nips-2008-One sketch for all: Theory and Application of Conditional Random Sampling