nips nips2005 nips2005-126 knowledge-graph by maker-knowledge-mining

126 nips-2005-Metric Learning by Collapsing Classes

Source: pdf

Author: Amir Globerson, Sam T. Roweis

Abstract: We present an algorithm for learning a quadratic Gaussian metric (Mahalanobis distance) for use in classiﬁcation tasks. Our method relies on the simple geometric intuition that a good metric is one under which points in the same class are simultaneously near each other and far from points in the other classes. We construct a convex optimization problem whose solution generates such a metric by trying to collapse all examples in the same class to a single point and push examples in other classes inﬁnitely far away. We show that when the metric we learn is used in simple classiﬁers, it yields substantial improvements over standard alternatives on a variety of problems. We also discuss how the learned metric may be used to obtain a compact low dimensional feature representation of the original input space, allowing more efﬁcient classiﬁcation with very little reduction in performance. 1 Supervised Learning of Metrics The problem of learning a distance measure (metric) over an input space is of fundamental importance in machine learning [10, 9], both supervised and unsupervised. When such measures are learned directly from the available data, they can be used to improve learning algorithms which rely on distance computations such as nearest neighbour classiﬁcation [5], supervised kernel machines (such as GPs or SVMs) and even unsupervised clustering algorithms [10]. Good similarity measures may also provide insight into the underlying structure of data (e.g. inter-protein distances), and may aid in building better data visualizations via embedding. In fact, there is a close link between distance learning and feature extraction since whenever we construct a feature ´Üµ for an input using a simple distance funcspace , we can measure distances between Ü½ Ü¾ ¾ tion (e.g. Euclidean) ´Ü½ µ ´Ü¾ µ℄ in feature space. Thus by ﬁxing , any feature extraction algorithm may be considered a metric learning method. Perhaps the simplest illustration of this approach is when the ´Üµ is a linear projection of Ü ¾ Ö so that ´Üµ Ï Ü. The Euclidean distance between ´Ü½ µ and ´Ü¾ µ is then the Mahalanobis distance ´Ü½ µ ´Ü¾ µ ¾ ´Ü½ Ü¾ µÌ ´Ü½ Ü¾ µ, where Ï Ì Ï is a positive semideﬁnite matrix. Much of the recent work on metric learning has indeed focused on learning Mahalanobis distances, i.e. learning the matrix . This is also the goal of the current work. A common approach to learning metrics is to assume some knowledge in the form of equiv- alence relations, i.e. which points should be close and which should be far (without specifying their exact distances). In the classiﬁcation setting there is a natural equivalence relation, namely whether two points are in the same class or not. One of the classical statistical methods which uses this idea for the Mahalanobis distance is Fisher’s Linear Discriminant Analysis (see e.g. [6]). Other more recent methods are [10, 9, 5] which seek to minimize various separation criteria between the classes under the new metric. In this work, we present a novel approach to learning such a metric. Our approach, the Maximally Collapsing Metric Learning algorithm (MCML), relies on the simple geometric intuition that if all points in the same class could be mapped into a single location in feature space and all points in other classes mapped to other locations, this would result in an ideal approximation of our equivalence relation. Our algorithm approximates this scenario via a stochastic selection rule, as in Neighborhood Component Analysis (NCA) [5]. However, unlike NCA, the optimization problem is convex and thus our method is completely speciﬁed by our objective function. Different initialization and optimization techniques may affect the speed of obtaining the solution but the ﬁnal solution itself is unique. We also show that our method approximates the local covariance structure of the data, as opposed to Linear Discriminant Analysis methods which use only global covariance structure. 2 The Approach of Collapsing Classes Given a set of Ò labeled examples ´Ü Ý µ, where Ü Ý ¾ ½ , we seek a space. We focus on Mahalanobis form metrics similarity measure between two points in ´Ü where Ü ´Ü µ ¾ Ü µÌ Ö and ´Ü Ü µ (1) is a positive semideﬁnite (PSD) matrix. Intuitively, what we want from a good metric is that it makes elements of in the same class look close whereas those in different classes appear far. Our approach starts with the ideal case when this is true in the most optimistic sense: same class points are at zero distance, and different class points are inﬁnitely far. Alternatively this can be viewed as mapping Ü via a linear projection Ï Ü ( Ï Ì Ï ), such that all points in the same class are mapped into the same point. This intuition is related to the analysis of spectral clustering [8], where the ideal case analysis of the algorithm results in all same cluster points being mapped to a single point. To learn a metric which approximates the ideal geometric setup described above, we introduce, for each training point, a conditional distribution over other points (as in [5]). such that Speciﬁcally, for each Ü we deﬁne a conditional distribution over points Ô ´ µ ½ È (2) If all points in the same class were mapped to a single point and inﬁnitely far from points in different classes, we would have the ideal “bi-level” distribution: Ô¼ ´ µ » Ò ½ ¼ Ý Ý Ý Ý (3) Furthermore, under very mild conditions, any set of points which achieves the above distribution must have the desired geometry. In particular, assume there are at least Ö · ¾ points in each class, where Ö rank ℄ (note that Ö Ö). Then Ô ´ µ Ô¼ ´ µ ( ) implies that under all points in the same class will be mapped to a single point, inﬁnitely far from other class points 1 . 1 Proof sketch: The inﬁnite separation between points of different classes follows simply from Thus it is natural to seek a matrix such that Ô ´ µ is as close as possible to Ô¼ ´ Since we are trying to match distributions, we minimize the KL divergence ÃÄ Ô¼ Ô℄: ÑÒ ÃÄ Ô¼ ´ µ Ô ´ µ℄ ¾ ÈË ×Ø µ. (4) The crucial property of this optimization problem is that it is convex in the matrix . To see this, ﬁrst note that any convex linear combination of feasible solutions « ¼ · ´½ «µ ½ s.t. ¼ « ½ is still a feasible solution, since the set of PSD matrices is convex. Next, we can show that ´ µ alway has a greater cost than either of the endpoints. To do this, we rewrite the objective function ´ µ ÃÄ Ô¼ ´ µ Ô´ µ℄ in the form 2 : È ´ µ ÐÓ Ý Ý Ô´ µ · Ý ÐÓ Ý where we assumed for simplicity that classes are equi-probable, yielding a multiplicative constant. To see why ´ µ is convex, ﬁrst note that ´Ü Ü µÌ ´Ü Ü µ is linear is a ÐÓ ÜÔ function of afﬁne functions of in , and thus convex. The function ÐÓ and is therefore also convex (see [4], page 74). È 2.1 Convex Duality Since our optimization problem is convex, it has an equivalent convex dual. Speciﬁcally, the convex dual of Eq. (4) is the following entropy maximization problem: À Ô´ Ñ Ü Ô´ µ where Ú Ü ×Ø µ℄ Ô¼ ´ Ú ÚÌ ℄ µ Ô´ µ Ú ÚÌ ℄ Ü , À ¡℄ is the entropy function and we require È Ô´ µ ¼ ½ (5) . To prove this duality we start with the proposed dual and obtain the original problem in Equation 4 as its dual. Write the Lagrangian for the above problem (where is PSD) 3 Ä´Ô ¬µ À ´Ô´ µµ Ì Ö´ ´ Ô¼ ´ Ú ÚÌ ℄ Ô Ú ÚÌ ℄µµµ ¬ ´ Ô´ µ ½µ The dual function is deﬁned as ´ ¬ µ Ñ ÒÔ Ä´Ô ¬ µ. To derive it, we ﬁrst solve for the minimizing Ô by setting the derivative of Ä´Ô ¬ µ w.r.t. Ô´ µ equal to zero. Ì ¼ ½ · ÐÓ Ô´ µ · Ì Ö ´ Ú Ú µ ¬ µ Ô´ µ ¬ ½ Ì Ö´ Ú ÚÌ µ Plugging solution È Ô thisThe dual to Ä Ô is¬ towe get ¬ . problem maximize È Ì Ö Ú Ú . ¬ , yielding ¬ ´ ´ µ ´ µ ½ Now note that Ì Ö´ ÐÓ Ú ÚÌ µ ÚÌ Ú ´ µ ´ Ìµ ´ µ Ì Ö´ ¬ µ. È Ô¼ Ú ÚÌ ℄µ · È¬ We can do this analytically w.r.t. , so we can write Ý Ý ÐÓ which is minus our original target function. Since ´ µ should be maximized, and we have the desired duality result (identifying with ). ¼ » µ ¼ when Ý Ý . For a given point Ü , all the points in its class satisfy Ô´ µ ½. Due to the structure of Ô´ µ in Equation 2, and because it is obeyed for all points in Ü¼ × class, this implies that all the points in that class are equidistant from each other. However, it is easy to show that the maximum number of different equidistant points (also known as the equilateral dimension [1]) in Ö dimensions is Ö · ½. Since by assumption we have at least Ö · ¾ points in the class of Ü , and maps points into Ö , it follows that all points are identical. 2 À Ô¼ ´ µ℄. Up to an additive constant 3 We consider the equivalent problem of minimizing minus entropy. Ô¼ ´ È 2.1.1 Relation to covariance based and embedding methods The convex dual derived above reveals an interesting relation to covariance based learning methods. The sufﬁcient statistics used by the algorithm are a set of Ò “spread” matrices. Each matrix is of the form Ô¼ ´ µ Ú ÚÌ ℄. The algorithm tries to ﬁnd a maximum entropy distribution which matches these matrices when averaged over the sample. This should be contrasted with the covariance matrices used in metric learning such as Fisher’s Discriminant Analysis. The latter uses the within and between class covariance matrices. The within covariance matrix is similar to the covariance matrix used here, but is calculated with respect to the class means, whereas here it is calculated separately for every point, and is centered on this point. This highlights the fact that MCML is not based on Gaussian assumptions where it is indeed sufﬁcient to calculate a single class covariance. Our method can also be thought of as a supervised version of the Stochastic Neighbour Embedding algorithm [7] in which the “target” distribution is Ô¼ (determined by the class labels) and the embedding points are not completely free but are instead constrained to be of the form Ï Ü . 2.2 Optimizing the Convex Objective Since the optimization problem in Equation 4 is convex, it is guaranteed to have only a single minimum which is the globally optimal solution4 . It can be optimized using any appropriate numerical convex optimization machinery; all methods will yield the same solution although some may be faster than others. One standard approach is to use interior point Newton methods. However, these algorithms require the Hessian to be calculated, which would require Ç´ µ resources, and could be prohibitive in our case. Instead, we have experimented with using a ﬁrst order gradient method, speciﬁcally the projected gradient approach as in [10]. At each iteration we take a small step in the direction of the negative gradient of the objective function5, followed by a projection back onto the PSD cone. This projection is performed simply by taking the eigen-decomposition of and removing the components with negative eigenvalues. The algorithm is summarized below: Ý µ, Ò Input: Set of labeled data points ´Ü Output: PSD metric which optimally collapses classes. ½ Initialization: Initialize ¼ to some PSD matrix (randomly or using some initialization heuristic). Iterate: È Ø ¯ ÔØ where Ü Ô Ü ¯ Calculate the eigen-decomposition of È È Ù ÙÌ , then set Ø Ø Ø ¯ Set Ø·½ ´ µ ´ ´ ¼´ µ µ ´ µµ´ µ´Ü Ü µÌ ·½ ·½ ·½ Ñ Ü´ ¼µÙ ÙÌ Of course in principle it is possible to optimize over the dual instead of the primal but in our case, if the training data consists of Ò points in Ö-dimensional space then the primal has only Ç´Ö¾ ¾µ variables while the dual has Ç´Ò¾ µ so it will almost always be more efﬁcient to operate on the primal directly. One exception to this case may be the kernel version (Section 4) where the primal is also of size Ç´Ò¾ µ. 4 When the data can be exactly collapsed into single class points, there will be multiple solutions at inﬁnity. However, this is very unlikely to happen in real data. 5 In the experiments, we used an Armijo like step size rule, as described in [3]. 3 Low Dimensional Projections for Feature Extraction The Mahalanobis distance under a metric can be interpreted as a linear projection of the original inputs by the square root of , followed by Euclidean distance in the projected space. Matrices which have less than full rank correspond to Mahalanobis distances based on low dimensional projections. Such metrics and the induced distances can be advantageous for several reasons [5]. First, low dimensional projections can substantially reduce the storage and computational requirements of a supervised method since only the projections of the training points must be stored and the manipulations at test time all occur in the lower dimensional feature space. Second, low dimensional projections re-represent the inputs, allowing for a supervised embedding or visualization of the original data. If we consider matrices with rank at most Õ , we can always represent them in the form Ï Ì Ï for some projection matrix Ï of size Õ ¢ Ö. This corresponds to projecting the original data into a Õ -dimensional space speciﬁed by the rows of Ï . However, rank constraints on a matrix are not convex [4], and hence the rank constrained problem is not convex and is likely to have local minima which make the optimization difﬁcult and illdeﬁned since it becomes sensitive to initial conditions and choice of optimization method. Luckily, there is an alternative approach to obtaining low dimensional projections, which does specify a unique solution by sequentially solving two globally tractable problems. This is the approach we follow here. First we solve for a (potentially) full rank metric using the convex program outlined above, and then obtain a low rank projection from it via spectral decomposition. This is done by diagonalizing into the form Ö Ú ÚÌ where ½ and Ú are the corre¾ Ö are eigenvalues of ½ sponding eigenvectors. To obtain a low rank projection we constrain the sum above to Õ include only the Õ terms corresponding to the Õ largest eigenvalues: Õ Ú ÚÌ . ½ The resulting projection is uniquely deﬁned (up to an irrelevant unitary transformation) as Ô Ì Ì Ï ´ ÚÕ ℄. ½ Õ µ Ú½ È È Ô In general, the projection returned by this approach is not guaranteed to be the same as the projection corresponding to minimizing our objective function subject to a rank constraint on unless the optimal metric is of rank less than or equal to Õ . However, as we show in the experimental results, it is often the case that for practical problems the optimal has an eigen-spectrum which is rapidly decaying, so that many of its eigenvalues are indeed very small, suggesting the low rank solution will be close to optimal. 4 Learning Metrics with Kernels It is interesting to consider the case where Ü are mapped into a high dimensional feature space ´Ü µ and a Mahalanobis distance is sought in this space. We focus on the case where dot products in the feature space may be expressed via a kernel function, such that ´Ü µ ¡ ´Ü µ ´Ü Ü µ for some kernel . We now show how our method can be changed to accommodate this setting, so that optimization depends only on dot products. Consider the regularized target function: Ê ´ µ ÃÄ Ô¼ ´ µ Ô´ µ℄ · Ì Ö´ µ (6) where the regularizing factor is equivalent to the Frobenius norm of the projection matrix Ï since Ì Ö´ µ Ï ¾. Deriving w.r.t. Ï we obtain Ï Í , where Í is some matrix which speciﬁes Ï as a linear combination of sample points, and the Ø row of the matrix Ì Í Ì Í . Deﬁning the PSD matrix is Ü . Thus is given by Í Ì Í , we can recast our optimization as looking for a PSD matrix , where the Mahalanobis distance is ´Ü Ü µÌ Ì ´Ü Ü µ ´ µÌ ´ µ, where we deﬁne Ü. This is exactly our original distance, with Ü replaced by , which depends only on dot products in space. The regularization term also depends solely on the dot products since Ì Ö´ µ Ì Ö´ Ì µ Ì Ö´ Ì µ Ì Ö´Ã µ, where Ã is the kernel matrix given Ì . Note that the trace is a linear function of , keeping the problem convex. by Ã Thus, as long as dot products can be represented via kernels, the optimization can be carried out without explicitly using the high dimensional space. To obtain a low dimensional solution, we follow the approach in Section 3: obtain a decomposition Î Ì Î 6 , and take the projection matrix to be the ﬁrst Õ rows of ¼ Î . Ì , and thus Ì Ì As a ﬁrst step, we calculate a matrix such that . Since is a correlation matrix for the rows of it can be shown (as in Kernel PCA) that its (left) eigenvectors are linear combinations of the rows of . Denoting by Î « the eigenvector matrix, we obtain, after some algebra, that « Ã Ì «. We conclude that « is an eigenvector of the matrix Ã Ì . Denote by « the matrix whose rows are orthonormal eigenvectors of Ã Ì . Then Î can be shown to be orthonormal if we set ¼ « . The ﬁnal projection will then be ¼ Î Ü « . Low dimensional Î projections will be obtained by keeping only the ﬁrst Õ components of this projection. 5 Experimental Results We compared our method to several metric learning algorithms on a supervised classiﬁcation task. Training data was ﬁrst used to learn a metric over the input space. Then this metric was used in a 1-nearest-neighbor algorithm to classify a test set. The datasets we investigated were taken from the UCI repository and have been used previously in evaluating supervised methods for metric learning [10, 5]. To these we added the USPS handwritten digits (downsampled to 8x8 pixels) and the YALE faces [2] (downsampled to 31x22). The algorithms used in the comparative evaluation were ¯ ¯ ¯ Fisher’s Linear Discriminant Analysis (LDA), which projects on the eigenvectors of ËÏ½ Ë where ËÏ Ë are the within and between class covariance matrices. The method of Xing et al [10] which minimizes the mean within class distance, while keeping the mean between class distance larger than one. Principal Component Analysis (PCA). There are several possibilities for scaling the PCA projections. We tested several, and report results of the empirically superior one (PCAW), which scales the projection components so that the covariance matrix after projection is the identity. PCAW often performs poorly on high dimensions, but globally outperforms all other variants. We also evaluated the kernel version of MCML with an RBF kernel (denoted by KMCML)7 . Since all methods allow projections to lower dimensions we compared performance for different projection dimensions 8 . The out-of sample performance results (based on 40 random splits of the data taking 70% for training and 30% for testing9 ) are shown in Figure 1. It can be seen that when used in a simple nearest-neighbour classiﬁer, the metric learned by MCML almost always performs as well as, or signiﬁcantly better than those learned by all other methods, across most dimensions. Furthermore, the kernel version of MCML outperforms the linear one on most datasets. Where Î is orthonormal, and the eigenvalues in are sorted in decreasing order. The regularization parameter and the width of the RBF kernel were chosen using 5 fold crossvalidation. KMCML was only evaluated for datasets with less than 1000 training points. 8 To obtain low dimensional mappings we used the approach outlined in Section 3. 9 Except for the larger datasets where 1000 random samples were used for training. 6 7 Wine 0.2 0.1 2 4 6 8 10 Projection Dimension 0.5 Error Rate Error Rate Balance Ion MCML PCAW LDA XING KMCML 0.3 0.25 0.4 0.2 0.15 0.3 0.1 0.2 0.05 10 20 30 Projection Dimension Soybean−small 0.1 1 2 3 Protein 4 Spam 0.6 0.4 0.25 0.5 0.4 0 5 10 15 0.2 0.3 0.2 0.15 20 5 10 Yale7 15 20 0.1 10 20 30 40 50 40 50 Digits Housing 0.6 0.4 0.4 0.3 0.4 0.35 0.2 0.3 0.2 0.1 0.25 10 20 30 40 50 5 10 15 10 20 30 Figure 1: Classiﬁcation error rate on several UCI datasets, USPS digits and YALE faces, for different projection dimensions. Algorithms are our Maximally Collapsing Metric Learning (MCML), Xing et.al.[10], PCA with whitening transformation (PCAW) and Fisher’s Discriminant Analysis (LDA). Standard errors of the means shown on curves. No results given for XING on YALE and KMCML on Digits and Spam due to the data size. 5.1 Comparison to non convex procedures The methods in the previous comparison are all well deﬁned, in the sense that they are not susceptible to local minima in the optimization. They also have the added advantage of obtaining projections to all dimensions using one optimization run. Below, we also compare the MCML results to the results of two non-convex procedures. The ﬁrst is the Non Convex variant of MCML (NMCML): The objective function of MCML can be optimized w.r.t the projection matrix Ï , where Ï Ì Ï . Although this is no longer a convex problem, it is not constrained and is thus easier to optimize. The second non convex method is Neighbourhood Components Analysis (NCA) [5], which attempts to directly minimize the error incurred by a nearest neighbor classiﬁer. For both methods we optimized the matrix Ï by restarting the optimization separately for each size of Ï . Minimization was performed using a conjugate gradient algorithm, initialized by LDA or randomly. Figure 2 shows results on a subset of the UCI datasets. It can be seen that the performance of NMCML is similar to that of MCML, although it is less stable, possibly due to local minima, and both methods usually outperform NCA. The inset in each ﬁgure shows the spectrum of the MCML matrix , revealing that it often drops quickly after a few dimensions. This illustrates the effectiveness of our two stage optimization procedure, and suggests its low dimensional solutions are close to optimal. 6 Discussion and Extensions We have presented an algorithm for learning maximally collapsing metrics (MCML), based on the intuition of collapsing classes into single points. MCML assumes that each class Balance Error Rate Wine MCML NMCML NCA 0.2 Soybean 0.4 0.1 0.3 0.1 0.05 0.2 0 2 4 6 8 10 Projection Dimension 0.1 1 0 2 Protein 3 4 5 10 Ion 15 20 Housing 0.4 0.6 0.3 0.5 0.35 0.2 0.4 0.3 0.3 0.25 5 10 15 20 0.1 10 20 30 5 10 15 Figure 2: Classiﬁcation error for non convex procedures, and the MCML method. Eigen-spectra for the MCML solution are shown in the inset. may be collapsed to a single point, at least approximately, and thus is only suitable for unimodal class distributions (or for simply connected sets if kernelization is used). However, if points belonging to a single class appear in several disconnected clusters in input (or feature) space, it is unlikely that MCML could collapse the class into a single point. It is possible that using a mixture of distributions, an EM-like algorithm can be constructed to accommodate this scenario. The method can also be used to learn low dimensional projections of the input space. We showed that it performs well, even across a range of projection dimensions, and consistently outperforms existing methods. Finally, we have shown how the method can be extended to projections in high dimensional feature spaces using the kernel trick. The resulting nonlinear method was shown to improve classiﬁcation results over the linear version. References Ò [1] N. Alon and P. Pudlak. Equilateral sets in ÐÔ . Geom. Funct. Anal., 13(3), 2003. [2] P. N. Belhumeur, J. Hespanha, and D. J. Kriegman. Eigenfaces vs. Fisherfaces: Recognition using class speciﬁc linear projection. In ECCV (1), 1996. [3] D.P. Bertsekas. On the Goldstein-Levitin-Polyak gradient projection method. IEEE Transaction on Automatic Control, 21(2):174–184, 1976. [4] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge Univ. Press, 2004. [5] J. Goldberger, S. Roweis, G. Hinton, and R. Salakhutdinov. Neighbourhood components analysis. In Advances in Neural Information Processing Systems (NIPS), 2004. [6] T. Hastie, R. Tibshirani, and J.H. Friedman. The elements of statistical learning: data mining, inference, and prediction. New York: Springer-Verlag, 2001. [7] G. Hinton and S. Roweis. Stochastic neighbor embedding. In Advances in Neural Information Processing Systems (NIPS), 2002. [8] A. Ng, M. Jordan, and Y. Weiss. On spectral clustering: Analysis and an algorithm. In Advances in Neural Information Processing Systems (NIPS), 2001. [9] N. Shental, T. Hertz, D. Weinshall, and M. Pavel. Adjustment learning and relevant component analysis. In Proc. of ECCV, 2002. [10] E. Xing, A. Ng, M. Jordan, and S. Russell. Distance metric learning, with application to clustering with side-information. In Advances in Neural Information Processing Systems (NIPS), 2004.

Reference: text

Summary: the most important sentenses genereted by tfidf model

sentIndex sentText sentNum sentScore

1 edu Abstract We present an algorithm for learning a quadratic Gaussian metric (Mahalanobis distance) for use in classiﬁcation tasks. [sent-6, score-0.323]

2 Our method relies on the simple geometric intuition that a good metric is one under which points in the same class are simultaneously near each other and far from points in the other classes. [sent-7, score-0.705]

3 We construct a convex optimization problem whose solution generates such a metric by trying to collapse all examples in the same class to a single point and push examples in other classes inﬁnitely far away. [sent-8, score-0.833]

4 We show that when the metric we learn is used in simple classiﬁers, it yields substantial improvements over standard alternatives on a variety of problems. [sent-9, score-0.292]

5 We also discuss how the learned metric may be used to obtain a compact low dimensional feature representation of the original input space, allowing more efﬁcient classiﬁcation with very little reduction in performance. [sent-10, score-0.496]

6 1 Supervised Learning of Metrics The problem of learning a distance measure (metric) over an input space is of fundamental importance in machine learning [10, 9], both supervised and unsupervised. [sent-11, score-0.228]

7 When such measures are learned directly from the available data, they can be used to improve learning algorithms which rely on distance computations such as nearest neighbour classiﬁcation [5], supervised kernel machines (such as GPs or SVMs) and even unsupervised clustering algorithms [10]. [sent-12, score-0.345]

8 In fact, there is a close link between distance learning and feature extraction since whenever we construct a feature ´Üµ for an input using a simple distance funcspace , we can measure distances between Ü½ Ü¾ ¾ tion (e. [sent-16, score-0.454]

9 Thus by ﬁxing , any feature extraction algorithm may be considered a metric learning method. [sent-19, score-0.423]

10 Perhaps the simplest illustration of this approach is when the ´Üµ is a linear projection of Ü ¾ Ö so that ´Üµ Ï Ü. [sent-20, score-0.254]

11 The Euclidean distance between ´Ü½ µ and ´Ü¾ µ is then the Mahalanobis distance ´Ü½ µ ´Ü¾ µ ¾ ´Ü½ Ü¾ µÌ ´Ü½ Ü¾ µ, where Ï Ì Ï is a positive semideﬁnite matrix. [sent-21, score-0.196]

12 Much of the recent work on metric learning has indeed focused on learning Mahalanobis distances, i. [sent-22, score-0.354]

13 which points should be close and which should be far (without specifying their exact distances). [sent-28, score-0.127]

14 In the classiﬁcation setting there is a natural equivalence relation, namely whether two points are in the same class or not. [sent-29, score-0.268]

15 However, unlike NCA, the optimization problem is convex and thus our method is completely speciﬁed by our objective function. [sent-37, score-0.324]

16 Different initialization and optimization techniques may affect the speed of obtaining the solution but the ﬁnal solution itself is unique. [sent-38, score-0.233]

17 We also show that our method approximates the local covariance structure of the data, as opposed to Linear Discriminant Analysis methods which use only global covariance structure. [sent-39, score-0.195]

18 We focus on Mahalanobis form metrics similarity measure between two points in ´Ü where Ü ´Ü µ ¾ Ü µÌ Ö and ´Ü Ü µ (1) is a positive semideﬁnite (PSD) matrix. [sent-41, score-0.217]

19 Intuitively, what we want from a good metric is that it makes elements of in the same class look close whereas those in different classes appear far. [sent-42, score-0.471]

20 Our approach starts with the ideal case when this is true in the most optimistic sense: same class points are at zero distance, and different class points are inﬁnitely far. [sent-43, score-0.539]

21 Alternatively this can be viewed as mapping Ü via a linear projection Ï Ü ( Ï Ì Ï ), such that all points in the same class are mapped into the same point. [sent-44, score-0.58]

22 This intuition is related to the analysis of spectral clustering [8], where the ideal case analysis of the algorithm results in all same cluster points being mapped to a single point. [sent-45, score-0.387]

23 To learn a metric which approximates the ideal geometric setup described above, we introduce, for each training point, a conditional distribution over other points (as in [5]). [sent-46, score-0.519]

24 In particular, assume there are at least Ö · ¾ points in each class, where Ö rank ℄ (note that Ö Ö). [sent-48, score-0.246]

25 Then Ô ´ µ Ô¼ ´ µ ( ) implies that under all points in the same class will be mapped to a single point, inﬁnitely far from other class points 1 . [sent-49, score-0.565]

26 (4) The crucial property of this optimization problem is that it is convex in the matrix . [sent-51, score-0.365]

27 To see this, ﬁrst note that any convex linear combination of feasible solutions « ¼ · ´½ «µ ½ s. [sent-52, score-0.232]

28 The function ÐÓ and is therefore also convex (see [4], page 74). [sent-58, score-0.202]

29 1 Convex Duality Since our optimization problem is convex, it has an equivalent convex dual. [sent-60, score-0.278]

30 To prove this duality we start with the proposed dual and obtain the original problem in Equation 4 as its dual. [sent-63, score-0.137]

31 Ì ¼ ½ · ÐÓ Ô´ µ · Ì Ö ´ Ú Ú µ ¬ µ Ô´ µ ¬ ½ Ì Ö´ Ú ÚÌ µ Plugging solution È Ô thisThe dual to Ä Ô is¬ towe get ¬ . [sent-69, score-0.123]

32 For a given point Ü , all the points in its class satisfy Ô´ µ ½. [sent-78, score-0.239]

33 Due to the structure of Ô´ µ in Equation 2, and because it is obeyed for all points in Ü¼ × class, this implies that all the points in that class are equidistant from each other. [sent-79, score-0.424]

34 However, it is easy to show that the maximum number of different equidistant points (also known as the equilateral dimension [1]) in Ö dimensions is Ö · ½. [sent-80, score-0.295]

35 Since by assumption we have at least Ö · ¾ points in the class of Ü , and maps points into Ö , it follows that all points are identical. [sent-81, score-0.493]

36 1 Relation to covariance based and embedding methods The convex dual derived above reveals an interesting relation to covariance based learning methods. [sent-86, score-0.522]

37 This should be contrasted with the covariance matrices used in metric learning such as Fisher’s Discriminant Analysis. [sent-90, score-0.438]

38 The latter uses the within and between class covariance matrices. [sent-91, score-0.19]

39 The within covariance matrix is similar to the covariance matrix used here, but is calculated with respect to the class means, whereas here it is calculated separately for every point, and is centered on this point. [sent-92, score-0.498]

40 This highlights the fact that MCML is not based on Gaussian assumptions where it is indeed sufﬁcient to calculate a single class covariance. [sent-93, score-0.141]

41 Our method can also be thought of as a supervised version of the Stochastic Neighbour Embedding algorithm [7] in which the “target” distribution is Ô¼ (determined by the class labels) and the embedding points are not completely free but are instead constrained to be of the form Ï Ü . [sent-94, score-0.361]

42 It can be optimized using any appropriate numerical convex optimization machinery; all methods will yield the same solution although some may be faster than others. [sent-97, score-0.351]

43 At each iteration we take a small step in the direction of the negative gradient of the objective function5, followed by a projection back onto the PSD cone. [sent-101, score-0.308]

44 This projection is performed simply by taking the eigen-decomposition of and removing the components with negative eigenvalues. [sent-102, score-0.224]

45 The algorithm is summarized below: Ý µ, Ò Input: Set of labeled data points ´Ü Output: PSD metric which optimally collapses classes. [sent-103, score-0.419]

46 ½ Initialization: Initialize ¼ to some PSD matrix (randomly or using some initialization heuristic). [sent-104, score-0.126]

47 One exception to this case may be the kernel version (Section 4) where the primal is also of size Ç´Ò¾ µ. [sent-106, score-0.132]

48 4 When the data can be exactly collapsed into single class points, there will be multiple solutions at inﬁnity. [sent-107, score-0.158]

49 3 Low Dimensional Projections for Feature Extraction The Mahalanobis distance under a metric can be interpreted as a linear projection of the original inputs by the square root of , followed by Euclidean distance in the projected space. [sent-110, score-0.742]

50 Matrices which have less than full rank correspond to Mahalanobis distances based on low dimensional projections. [sent-111, score-0.338]

51 Such metrics and the induced distances can be advantageous for several reasons [5]. [sent-112, score-0.161]

52 First, low dimensional projections can substantially reduce the storage and computational requirements of a supervised method since only the projections of the training points must be stored and the manipulations at test time all occur in the lower dimensional feature space. [sent-113, score-0.749]

53 Second, low dimensional projections re-represent the inputs, allowing for a supervised embedding or visualization of the original data. [sent-114, score-0.397]

54 If we consider matrices with rank at most Õ , we can always represent them in the form Ï Ì Ï for some projection matrix Ï of size Õ ¢ Ö. [sent-115, score-0.467]

55 However, rank constraints on a matrix are not convex [4], and hence the rank constrained problem is not convex and is likely to have local minima which make the optimization difﬁcult and illdeﬁned since it becomes sensitive to initial conditions and choice of optimization method. [sent-117, score-0.92]

56 Luckily, there is an alternative approach to obtaining low dimensional projections, which does specify a unique solution by sequentially solving two globally tractable problems. [sent-118, score-0.262]

57 First we solve for a (potentially) full rank metric using the convex program outlined above, and then obtain a low rank projection from it via spectral decomposition. [sent-120, score-1.073]

58 To obtain a low rank projection we constrain the sum above to Õ include only the Õ terms corresponding to the Õ largest eigenvalues: Õ Ú ÚÌ . [sent-122, score-0.395]

59 ½ The resulting projection is uniquely deﬁned (up to an irrelevant unitary transformation) as Ô Ì Ì Ï ´ ÚÕ ℄. [sent-123, score-0.224]

60 ½ Õ µ Ú½ È È Ô In general, the projection returned by this approach is not guaranteed to be the same as the projection corresponding to minimizing our objective function subject to a rank constraint on unless the optimal metric is of rank less than or equal to Õ . [sent-124, score-1.024]

61 However, as we show in the experimental results, it is often the case that for practical problems the optimal has an eigen-spectrum which is rapidly decaying, so that many of its eigenvalues are indeed very small, suggesting the low rank solution will be close to optimal. [sent-125, score-0.256]

62 4 Learning Metrics with Kernels It is interesting to consider the case where Ü are mapped into a high dimensional feature space ´Ü µ and a Mahalanobis distance is sought in this space. [sent-126, score-0.337]

63 We focus on the case where dot products in the feature space may be expressed via a kernel function, such that ´Ü µ ¡ ´Ü µ ´Ü Ü µ for some kernel . [sent-127, score-0.301]

64 We now show how our method can be changed to accommodate this setting, so that optimization depends only on dot products. [sent-128, score-0.179]

65 Consider the regularized target function: Ê ´ µ ÃÄ Ô¼ ´ µ Ô´ µ℄ · Ì Ö´ µ (6) where the regularizing factor is equivalent to the Frobenius norm of the projection matrix Ï since Ì Ö´ µ Ï ¾. [sent-129, score-0.311]

66 Ï we obtain Ï Í , where Í is some matrix which speciﬁes Ï as a linear combination of sample points, and the Ø row of the matrix Ì Í Ì Í . [sent-133, score-0.204]

67 Thus is given by Í Ì Í , we can recast our optimization as looking for a PSD matrix , where the Mahalanobis distance is ´Ü Ü µÌ Ì ´Ü Ü µ ´ µÌ ´ µ, where we deﬁne Ü. [sent-135, score-0.261]

68 This is exactly our original distance, with Ü replaced by , which depends only on dot products in space. [sent-136, score-0.123]

69 The regularization term also depends solely on the dot products since Ì Ö´ µ Ì Ö´ Ì µ Ì Ö´ Ì µ Ì Ö´Ã µ, where Ã is the kernel matrix given Ì . [sent-137, score-0.271]

70 by Ã Thus, as long as dot products can be represented via kernels, the optimization can be carried out without explicitly using the high dimensional space. [sent-139, score-0.295]

71 To obtain a low dimensional solution, we follow the approach in Section 3: obtain a decomposition Î Ì Î 6 , and take the projection matrix to be the ﬁrst Õ rows of ¼ Î . [sent-140, score-0.509]

72 Since is a correlation matrix for the rows of it can be shown (as in Kernel PCA) that its (left) eigenvectors are linear combinations of the rows of . [sent-142, score-0.251]

73 Denote by « the matrix whose rows are orthonormal eigenvectors of Ã Ì . [sent-145, score-0.215]

74 Low dimensional Î projections will be obtained by keeping only the ﬁrst Õ components of this projection. [sent-148, score-0.262]

75 5 Experimental Results We compared our method to several metric learning algorithms on a supervised classiﬁcation task. [sent-149, score-0.391]

76 Training data was ﬁrst used to learn a metric over the input space. [sent-150, score-0.292]

77 Then this metric was used in a 1-nearest-neighbor algorithm to classify a test set. [sent-151, score-0.292]

78 The datasets we investigated were taken from the UCI repository and have been used previously in evaluating supervised methods for metric learning [10, 5]. [sent-152, score-0.429]

79 The algorithms used in the comparative evaluation were ¯ ¯ ¯ Fisher’s Linear Discriminant Analysis (LDA), which projects on the eigenvectors of ËÏ½ Ë where ËÏ Ë are the within and between class covariance matrices. [sent-154, score-0.224]

80 The method of Xing et al [10] which minimizes the mean within class distance, while keeping the mean between class distance larger than one. [sent-155, score-0.361]

81 We tested several, and report results of the empirically superior one (PCAW), which scales the projection components so that the covariance matrix after projection is the identity. [sent-158, score-0.613]

82 We also evaluated the kernel version of MCML with an RBF kernel (denoted by KMCML)7 . [sent-160, score-0.122]

83 Since all methods allow projections to lower dimensions we compared performance for different projection dimensions 8 . [sent-161, score-0.455]

84 It can be seen that when used in a simple nearest-neighbour classiﬁer, the metric learned by MCML almost always performs as well as, or signiﬁcantly better than those learned by all other methods, across most dimensions. [sent-163, score-0.292]

85 8 To obtain low dimensional mappings we used the approach outlined in Section 3. [sent-168, score-0.185]

86 25 10 20 30 40 50 5 10 15 10 20 30 Figure 1: Classiﬁcation error rate on several UCI datasets, USPS digits and YALE faces, for different projection dimensions. [sent-204, score-0.282]

87 1 Comparison to non convex procedures The methods in the previous comparison are all well deﬁned, in the sense that they are not susceptible to local minima in the optimization. [sent-211, score-0.328]

88 They also have the added advantage of obtaining projections to all dimensions using one optimization run. [sent-212, score-0.285]

89 Although this is no longer a convex problem, it is not constrained and is thus easier to optimize. [sent-217, score-0.202]

90 The second non convex method is Neighbourhood Components Analysis (NCA) [5], which attempts to directly minimize the error incurred by a nearest neighbor classiﬁer. [sent-218, score-0.26]

91 For both methods we optimized the matrix Ï by restarting the optimization separately for each size of Ï . [sent-219, score-0.192]

92 This illustrates the effectiveness of our two stage optimization procedure, and suggests its low dimensional solutions are close to optimal. [sent-224, score-0.224]

93 6 Discussion and Extensions We have presented an algorithm for learning maximally collapsing metrics (MCML), based on the intuition of collapsing classes into single points. [sent-225, score-0.623]

94 1 10 20 30 5 10 15 Figure 2: Classiﬁcation error for non convex procedures, and the MCML method. [sent-245, score-0.26]

95 may be collapsed to a single point, at least approximately, and thus is only suitable for unimodal class distributions (or for simply connected sets if kernelization is used). [sent-247, score-0.158]

96 However, if points belonging to a single class appear in several disconnected clusters in input (or feature) space, it is unlikely that MCML could collapse the class into a single point. [sent-248, score-0.42]

97 The method can also be used to learn low dimensional projections of the input space. [sent-250, score-0.275]

98 We showed that it performs well, even across a range of projection dimensions, and consistently outperforms existing methods. [sent-251, score-0.253]

99 Finally, we have shown how the method can be extended to projections in high dimensional feature spaces using the kernel trick. [sent-252, score-0.34]

100 Distance metric learning, with application to clustering with side-information. [sent-320, score-0.329]

similar papers computed by tfidf model

tfidf for this paper:

wordName wordTfidf (topN-words)

[('mcml', 0.522), ('metric', 0.292), ('projection', 0.224), ('mahalanobis', 0.207), ('psd', 0.202), ('convex', 0.202), ('collapsing', 0.174), ('points', 0.127), ('projections', 0.127), ('rank', 0.119), ('kmcml', 0.116), ('pcaw', 0.116), ('class', 0.112), ('nca', 0.101), ('xing', 0.099), ('distance', 0.098), ('dimensional', 0.096), ('metrics', 0.09), ('matrix', 0.087), ('nmcml', 0.087), ('mapped', 0.087), ('discriminant', 0.08), ('dual', 0.079), ('covariance', 0.078), ('optimization', 0.076), ('distances', 0.071), ('dot', 0.071), ('primal', 0.071), ('lda', 0.068), ('supervised', 0.068), ('classes', 0.067), ('kernel', 0.061), ('ideal', 0.061), ('yale', 0.061), ('fisher', 0.06), ('non', 0.058), ('digits', 0.058), ('equidistant', 0.058), ('equilateral', 0.058), ('soybean', 0.058), ('spam', 0.058), ('duality', 0.058), ('feature', 0.056), ('embedding', 0.054), ('nitely', 0.053), ('low', 0.052), ('products', 0.052), ('dimensions', 0.052), ('neighbour', 0.05), ('rows', 0.05), ('intuition', 0.047), ('uci', 0.046), ('objective', 0.046), ('collapsed', 0.046), ('wine', 0.046), ('ion', 0.046), ('classi', 0.045), ('pca', 0.045), ('solution', 0.044), ('orthonormal', 0.044), ('extraction', 0.044), ('neighbourhood', 0.043), ('seek', 0.042), ('eigenvalues', 0.041), ('euclidean', 0.041), ('collapse', 0.04), ('downsampled', 0.04), ('housing', 0.04), ('maximally', 0.04), ('roweis', 0.04), ('globally', 0.04), ('keeping', 0.039), ('initialization', 0.039), ('minima', 0.039), ('approximates', 0.039), ('eccv', 0.038), ('gradient', 0.038), ('datasets', 0.038), ('matrices', 0.037), ('clustering', 0.037), ('minus', 0.037), ('outlined', 0.037), ('usps', 0.037), ('eigenvectors', 0.034), ('protein', 0.032), ('accommodate', 0.032), ('learning', 0.031), ('entropy', 0.03), ('obtaining', 0.03), ('linear', 0.03), ('optimized', 0.029), ('unlikely', 0.029), ('procedures', 0.029), ('equivalence', 0.029), ('calculate', 0.029), ('outperforms', 0.029), ('hinton', 0.028), ('spectral', 0.028), ('calculated', 0.028), ('semide', 0.028)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 1.0 126 nips-2005-Metric Learning by Collapsing Classes

Author: Amir Globerson, Sam T. Roweis

2 0.26965424 57 nips-2005-Distance Metric Learning for Large Margin Nearest Neighbor Classification

Author: Kilian Q. Weinberger, John Blitzer, Lawrence K. Saul

Abstract: We show how to learn a Mahanalobis distance metric for k-nearest neighbor (kNN) classiﬁcation by semideﬁnite programming. The metric is trained with the goal that the k-nearest neighbors always belong to the same class while examples from different classes are separated by a large margin. On seven data sets of varying size and difﬁculty, we ﬁnd that metrics trained in this way lead to signiﬁcant improvements in kNN classiﬁcation—for example, achieving a test error rate of 1.3% on the MNIST handwritten digits. As in support vector machines (SVMs), the learning problem reduces to a convex optimization based on the hinge loss. Unlike learning in SVMs, however, our framework requires no modiﬁcation or extension for problems in multiway (as opposed to binary) classiﬁcation. 1

3 0.1534079 166 nips-2005-Robust Fisher Discriminant Analysis

Author: Seung-jean Kim, Alessandro Magnani, Stephen Boyd

Abstract: Fisher linear discriminant analysis (LDA) can be sensitive to the problem data. Robust Fisher LDA can systematically alleviate the sensitivity problem by explicitly incorporating a model of data uncertainty in a classiﬁcation problem and optimizing for the worst-case scenario under this model. The main contribution of this paper is show that with general convex uncertainty models on the problem data, robust Fisher LDA can be carried out using convex optimization. For a certain type of product form uncertainty model, robust Fisher LDA can be carried out at a cost comparable to standard Fisher LDA. The method is demonstrated with some numerical examples. Finally, we show how to extend these results to robust kernel Fisher discriminant analysis, i.e., robust Fisher LDA in a high dimensional feature space. 1

4 0.15057719 42 nips-2005-Combining Graph Laplacians for Semi--Supervised Learning

Author: Andreas Argyriou, Mark Herbster, Massimiliano Pontil

Abstract: A foundational problem in semi-supervised learning is the construction of a graph underlying the data. We propose to use a method which optimally combines a number of differently constructed graphs. For each of these graphs we associate a basic graph kernel. We then compute an optimal combined kernel. This kernel solves an extended regularization problem which requires a joint minimization over both the data and the set of graph kernels. We present encouraging results on different OCR tasks where the optimal combined kernel is computed from graphs constructed with a variety of distances functions and the ‘k’ in nearest neighbors. 1

5 0.11411314 50 nips-2005-Convex Neural Networks

Author: Yoshua Bengio, Nicolas L. Roux, Pascal Vincent, Olivier Delalleau, Patrice Marcotte

Abstract: Convexity has recently received a lot of attention in the machine learning community, and the lack of convexity has been seen as a major disadvantage of many learning algorithms, such as multi-layer artiﬁcial neural networks. We show that training multi-layer neural networks in which the number of hidden units is learned can be viewed as a convex optimization problem. This problem involves an inﬁnite number of variables, but can be solved by incrementally inserting a hidden unit at a time, each time ﬁnding a linear classiﬁer that minimizes a weighted sum of errors. 1

6 0.11058544 147 nips-2005-On the Convergence of Eigenspaces in Kernel Principal Component Analysis

7 0.10604872 27 nips-2005-Analysis of Spectral Kernel Design based Semi-supervised Learning

8 0.096709117 82 nips-2005-Generalization Error Bounds for Aggregation by Mirror Descent with Averaging

9 0.091895349 129 nips-2005-Modeling Neural Population Spiking Activity with Gibbs Distributions

10 0.086955577 202 nips-2005-Variational EM Algorithms for Non-Gaussian Latent Variable Models

11 0.086754173 56 nips-2005-Diffusion Maps, Spectral Clustering and Eigenfunctions of Fokker-Planck Operators

12 0.079850078 10 nips-2005-A General and Efficient Multiple Kernel Learning Algorithm

13 0.079496756 79 nips-2005-Fusion of Similarity Data in Clustering

14 0.076403998 123 nips-2005-Maximum Margin Semi-Supervised Learning for Structured Variables

15 0.07076709 58 nips-2005-Divergences, surrogate loss functions and experimental design

16 0.070129588 92 nips-2005-Hyperparameter and Kernel Learning for Graph Based Semi-Supervised Classification

17 0.069838852 102 nips-2005-Kernelized Infomax Clustering

18 0.068157606 114 nips-2005-Learning Rankings via Convex Hull Separation

19 0.064111941 137 nips-2005-Non-Gaussian Component Analysis: a Semi-parametric Framework for Linear Dimension Reduction

20 0.062594362 18 nips-2005-Active Learning For Identifying Function Threshold Boundaries

similar papers computed by lsi model

lsi for this paper:

topicId topicWeight

[(0, 0.246), (1, 0.121), (2, -0.091), (3, -0.038), (4, -0.035), (5, 0.054), (6, 0.093), (7, -0.075), (8, 0.009), (9, -0.064), (10, 0.046), (11, -0.122), (12, -0.116), (13, 0.021), (14, -0.008), (15, -0.057), (16, -0.102), (17, -0.078), (18, -0.311), (19, -0.05), (20, -0.055), (21, -0.019), (22, -0.021), (23, 0.158), (24, -0.169), (25, 0.047), (26, 0.098), (27, -0.058), (28, -0.048), (29, -0.028), (30, 0.062), (31, 0.003), (32, 0.131), (33, 0.04), (34, 0.056), (35, 0.004), (36, -0.03), (37, 0.018), (38, -0.061), (39, 0.068), (40, 0.043), (41, -0.055), (42, -0.119), (43, 0.171), (44, -0.069), (45, 0.04), (46, -0.002), (47, -0.026), (48, 0.075), (49, 0.024)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.93948579 126 nips-2005-Metric Learning by Collapsing Classes

Author: Amir Globerson, Sam T. Roweis

2 0.70128936 57 nips-2005-Distance Metric Learning for Large Margin Nearest Neighbor Classification

Author: Kilian Q. Weinberger, John Blitzer, Lawrence K. Saul

3 0.62617671 166 nips-2005-Robust Fisher Discriminant Analysis

Author: Seung-jean Kim, Alessandro Magnani, Stephen Boyd

4 0.52356452 189 nips-2005-Tensor Subspace Analysis

Author: Xiaofei He, Deng Cai, Partha Niyogi

Abstract: Previous work has demonstrated that the image variations of many objects (human faces in particular) under variable lighting can be effectively modeled by low dimensional linear spaces. The typical linear subspace learning algorithms include Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and Locality Preserving Projection (LPP). All of these methods consider an n1 × n2 image as a high dimensional vector in Rn1 ×n2 , while an image represented in the plane is intrinsically a matrix. In this paper, we propose a new algorithm called Tensor Subspace Analysis (TSA). TSA considers an image as the second order tensor in Rn1 ⊗ Rn2 , where Rn1 and Rn2 are two vector spaces. The relationship between the column vectors of the image matrix and that between the row vectors can be naturally characterized by TSA. TSA detects the intrinsic local geometrical structure of the tensor space by learning a lower dimensional tensor subspace. We compare our proposed approach with PCA, LDA and LPP methods on two standard databases. Experimental results demonstrate that TSA achieves better recognition rate, while being much more efﬁcient. 1

5 0.51307708 42 nips-2005-Combining Graph Laplacians for Semi--Supervised Learning

Author: Andreas Argyriou, Mark Herbster, Massimiliano Pontil

6 0.4455854 180 nips-2005-Spectral Bounds for Sparse PCA: Exact and Greedy Algorithms

7 0.41249225 159 nips-2005-Q-Clustering

8 0.41003504 82 nips-2005-Generalization Error Bounds for Aggregation by Mirror Descent with Averaging

9 0.40950659 10 nips-2005-A General and Efficient Multiple Kernel Learning Algorithm

10 0.40723491 71 nips-2005-Fast Krylov Methods for N-Body Learning

11 0.40472558 137 nips-2005-Non-Gaussian Component Analysis: a Semi-parametric Framework for Linear Dimension Reduction

12 0.4036535 43 nips-2005-Comparing the Effects of Different Weight Distributions on Finding Sparse Representations

13 0.40117356 18 nips-2005-Active Learning For Identifying Function Threshold Boundaries

14 0.3902975 50 nips-2005-Convex Neural Networks

15 0.38882494 79 nips-2005-Fusion of Similarity Data in Clustering

16 0.38766131 62 nips-2005-Efficient Estimation of OOMs

17 0.38708514 34 nips-2005-Bayesian Surprise Attracts Human Attention

18 0.38504651 104 nips-2005-Laplacian Score for Feature Selection

19 0.38435048 132 nips-2005-Nearest Neighbor Based Feature Selection for Regression and its Application to Neural Activity

20 0.37157953 58 nips-2005-Divergences, surrogate loss functions and experimental design

similar papers computed by lda model

lda for this paper:

topicId topicWeight

[(3, 0.052), (10, 0.032), (27, 0.035), (31, 0.028), (34, 0.155), (50, 0.341), (55, 0.029), (69, 0.053), (73, 0.058), (77, 0.011), (88, 0.086), (91, 0.033)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.83617032 126 nips-2005-Metric Learning by Collapsing Classes

Author: Amir Globerson, Sam T. Roweis

2 0.78460622 123 nips-2005-Maximum Margin Semi-Supervised Learning for Structured Variables

Author: Y. Altun, D. McAllester, M. Belkin

Abstract: Many real-world classiﬁcation problems involve the prediction of multiple inter-dependent variables forming some structural dependency. Recent progress in machine learning has mainly focused on supervised classiﬁcation of such structured variables. In this paper, we investigate structured classiﬁcation in a semi-supervised setting. We present a discriminative approach that utilizes the intrinsic geometry of input patterns revealed by unlabeled data points and we derive a maximum-margin formulation of semi-supervised learning for structured variables. Unlike transductive algorithms, our formulation naturally extends to new test points. 1

3 0.7170943 129 nips-2005-Modeling Neural Population Spiking Activity with Gibbs Distributions

Author: Frank Wood, Stefan Roth, Michael J. Black

Abstract: Probabilistic modeling of correlated neural population ﬁring activity is central to understanding the neural code and building practical decoding algorithms. No parametric models currently exist for modeling multivariate correlated neural data and the high dimensional nature of the data makes fully non-parametric methods impractical. To address these problems we propose an energy-based model in which the joint probability of neural activity is represented using learned functions of the 1D marginal histograms of the data. The parameters of the model are learned using contrastive divergence and an optimization procedure for ﬁnding appropriate marginal directions. We evaluate the method using real data recorded from a population of motor cortical neurons. In particular, we model the joint probability of population spiking times and 2D hand position and show that the likelihood of test data under our model is signiﬁcantly higher than under other models. These results suggest that our model captures correlations in the ﬁring activity. Our rich probabilistic model of neural population activity is a step towards both measurement of the importance of correlations in neural coding and improved decoding of population activity. 1

4 0.60097742 57 nips-2005-Distance Metric Learning for Large Margin Nearest Neighbor Classification

Author: Kilian Q. Weinberger, John Blitzer, Lawrence K. Saul

5 0.57034618 184 nips-2005-Structured Prediction via the Extragradient Method

Author: Ben Taskar, Simon Lacoste-Julian, Michael I. Jordan

Abstract: We present a simple and scalable algorithm for large-margin estimation of structured models, including an important class of Markov networks and combinatorial models. We formulate the estimation problem as a convex-concave saddle-point problem and apply the extragradient method, yielding an algorithm with linear convergence using simple gradient and projection calculations. The projection step can be solved using combinatorial algorithms for min-cost quadratic ﬂow. This makes the approach an efﬁcient alternative to formulations based on reductions to a quadratic program (QP). We present experiments on two very different structured prediction tasks: 3D image segmentation and word alignment, illustrating the favorable scaling properties of our algorithm. 1

6 0.55227542 114 nips-2005-Learning Rankings via Convex Hull Separation

7 0.53155303 190 nips-2005-The Curse of Highly Variable Functions for Local Kernel Machines

8 0.53151226 92 nips-2005-Hyperparameter and Kernel Learning for Graph Based Semi-Supervised Classification

9 0.52422059 137 nips-2005-Non-Gaussian Component Analysis: a Semi-parametric Framework for Linear Dimension Reduction

10 0.52168882 50 nips-2005-Convex Neural Networks

11 0.51553553 177 nips-2005-Size Regularized Cut for Data Clustering

12 0.51183045 102 nips-2005-Kernelized Infomax Clustering

13 0.51114082 189 nips-2005-Tensor Subspace Analysis

14 0.50881916 56 nips-2005-Diffusion Maps, Spectral Clustering and Eigenfunctions of Fokker-Planck Operators

15 0.50844646 58 nips-2005-Divergences, surrogate loss functions and experimental design

16 0.5080992 43 nips-2005-Comparing the Effects of Different Weight Distributions on Finding Sparse Representations

17 0.50235444 31 nips-2005-Asymptotics of Gaussian Regularized Least Squares

18 0.50141776 77 nips-2005-From Lasso regression to Feature vector machine

19 0.50061297 105 nips-2005-Large-Scale Multiclass Transduction

20 0.49906045 143 nips-2005-Off-Road Obstacle Avoidance through End-to-End Learning