nips nips2002 nips2002-49 knowledge-graph by maker-knowledge-mining

49 nips-2002-Charting a Manifold

Source: pdf

Author: Matthew Brand

Abstract: We construct a nonlinear mapping from a high-dimensional sample space to a low-dimensional vector space, effectively recovering a Cartesian coordinate system for the manifold from which the data is sampled. The mapping preserves local geometric relations in the manifold and is pseudo-invertible. We show how to estimate the intrinsic dimensionality of the manifold from samples, decompose the sample data into locally linear low-dimensional patches, merge these patches into a single lowdimensional coordinate system, and compute forward and reverse mappings between the sample and coordinate spaces. The objective functions are convex and their solutions are given in closed form. 1 Nonlinear dimensionality reduction (NLDR) by charting Charting is the problem of assigning a low-dimensional coordinate system to data points in a high-dimensional sample space. It is presumed that the data lies on or near a lowdimensional manifold embedded in the sample space, and that there exists a 1-to-1 smooth nonlinear transform between the manifold and a low-dimensional vector space. The datamodeler’s goal is to estimate smooth continuous mappings between the sample and coordinate spaces. Often this analysis will shed light on the intrinsic variables of the datagenerating phenomenon, for example, revealing perceptual or conﬁguration spaces. Our goal is to ﬁnd a mapping—expressed as a kernel-based mixture of linear projections— that minimizes information loss about the density and relative locations of sample points. This constraint is expressed in a posterior that combines a standard gaussian mixture model (GMM) likelihood function with a prior that penalizes uncertainty due to inconsistent projections in the mixture. Section 3 develops a special case where this posterior is unimodal and maximizable in closed form, yielding a GMM whose covariances reveal a patchwork of overlapping locally linear subspaces that cover the manifold. Section 4 shows that for this (or any) GMM and a choice of reduced dimension d, there is a unique, closed-form solution for a minimally distorting merger of the subspaces into a d-dimensional coordinate space, as well as an reverse mapping deﬁning the surface of the manifold in the sample space. The intrinsic dimensionality d of the data manifold can be estimated from the growth process of point-to-point distances. In analogy to differential geometry, we call the subspaces “charts” and their merger the “connection.” Section 5 considers example problems where these methods are used to untie knots, unroll and untwist sheets, and visualize video data. 1.1 Background Topology-neutral NLDR algorithms can be divided into those that compute mappings, and those that directly compute low-dimensional embeddings. The ﬁ has its roots in mapeld ping algorithms: DeMers and Cottrell [3] proposed using auto-encoding neural networks with a hidden layer “ bottleneck,” effectively casting dimensionality reduction as a compression problem. Hastie deﬁ principal curves [5] as nonparametric 1 D curves that pass ned through the center of “ nearby” data points. A rich literature has grown up around properly regularizing this approach and extending it to surfaces. Smola and colleagues [10] analyzed the NLDR problem in the broader framework of regularized quantization methods. More recent advances aim for embeddings: Gomes and Mojsilovic [4] treat manifold completion as an anisotropic diffusion problem, iteratively expanding points until they connect to their neighbors. The I SO M AP algorithm [12] represents remote distances as sums of a trusted set of distances between immediate neighbors, then uses multidimensional scaling to compute a low-dimensional embedding that minimally distorts all distances. The locally linear embedding algorithm (LLE) [9] represents each point as a weighted combination of a trusted set of nearest neighbors, then computes a minimally distorting low-dimensional barycentric embedding. They have complementary strengths: I SO M AP handles holes well but can fail if the data hull is nonconvex [12]; and vice versa for LLE [9]. Both offer embeddings without mappings. It has been noted that trusted-set methods are vulnerable to noise because they consider the subset of point-to-point relationships that has the lowest signal-to-noise ratio; small changes to the trusted set can induce large changes in the set of constraints on the embedding, making solutions unstable [1]. In a return to mapping, Roweis and colleagues [8] proposed global coordination— learning a mixture of locally linear projections from sample to coordinate space. They constructed a posterior that penalizes distortions in the mapping, and gave a expectation-maximization (EM) training rule. Innovative use of variational methods highlighted the difﬁ culty of even hill-climbing their multimodal posterior. Like [2, 7, 6, 8], the method we develop below is a decomposition of the manifold into locally linear neighborhoods. It bears closest relation to global coordination [8], although by a different construction of the problem, we avoid hill-climbing a spiky posterior and instead develop a closed-form solution. 2 Estimating locally linear scale and intrinsic dimensionality . We begin with matrix of sample points Y = [y1 , · · · , yN ], yn ∈ RD populating a Ddimensional sample space, and a conjecture that these points are samples from a manifold M of intrinsic dimensionality d < D. We seek a mapping onto a vector space . G(Y) → X = [x1 , · · · , xN ], xn ∈ Rd and 1-to-1 reverse mapping G−1 (X) → Y such that local relations between nearby points are preserved (this will be formalized below). The map G should be non-catastrophic, that is, without folds: Parallel lines on the manifold in RD should map to continuous smooth non-intersecting curves in Rd . This guarantees that linear operations on X such as interpolation will have reasonable analogues on Y. Smoothness means that at some scale r the mapping from a neighborhood on M to Rd is effectively linear. Consider a ball of radius r centered on a data point and containing n(r) data points. The count n(r) grows as rd , but only at the locally linear scale; the grow rate is inﬂated by isotropic noise at smaller scales and by embedding curvature at larger scales. . To estimate r, we look at how the r-ball grows as points are added to it, tracking c(r) = d d log n(r) log r. At noise scales, c(r) ≈ 1/D < 1/d, because noise has distributed points in all directions with equal probability. At the scale at which curvature becomes signiﬁ cant, c(r) < 1/d, because the manifold is no longer perpendicular to the surface of the ball, so the ball does not have to grow as fast to accommodate new points. At the locally linear scale, the process peaks at c(r) = 1/d, because points are distributed only in the directions of the manifold’s local tangent space. The maximum of c(r) therefore gives an estimate of both the scale and the local dimensionality of the manifold (see ﬁ gure 1), provided that the ball hasn’t expanded to a manifold boundary— boundaries have lower dimension than Scale behavior of a 1D manifold in 2-space Point−count growth process on a 2D manifold in 3−space 1 10 radial growth process 1D hypothesis 2D hypothesis 3D hypothesis radius (log scale) samples noise scale locally linear scale curvature scale 0 10 2 1 10 2 10 #points (log scale) 3 10 Figure 1: Point growth processes. L EFT: At the locally linear scale, the number of points in an r-ball grows as rd ; at noise and curvature scales it grows faster. R IGHT: Using the point-count growth process to ﬁ the intrinsic dimensionality of a 2D manifold nonlinearly nd embedded in 3-space (see ﬁ gure 2). Lines of slope 1/3 , 1/2 , and 1 are ﬁ tted to sections of the log r/ log nr curve. For neighborhoods of radius r ≈ 1 with roughly n ≈ 10 points, the slope peaks at 1/2 indicating a dimensionality of d = 2. Below that, the data appears 3 D because it is dominated by noise (except for n ≤ D points); above, the data appears >2 D because of manifold curvature. As the r-ball expands to cover the entire data-set the dimensionality appears to drop to 1 as the process begins to track the 1D edges of the 2D sheet. the manifold. For low-dimensional manifolds such as sheets, the boundary submanifolds (edges and corners) are very small relative to the full manifold, so the boundary effect is typically limited to a small rise in c(r) as r approaches the scale of the entire data set. In practice, our code simply expands an r-ball at every point and looks for the ﬁ peak in rst c(r), averaged over many nearby r-balls. One can estimate d and r globally or per-point. 3 Charting the data In the charting step we ﬁ a soft partitioning of the data into locally linear low-dimensional nd neighborhoods, as a prelude to computing the connection that gives the global lowdimensional embedding. To minimize information loss in the connection, we require that the data points project into a subspace associated with each neighborhood with (1) minimal loss of local variance and (2) maximal agreement of the projections of nearby points into nearby neighborhoods. Criterion (1) is served by maximizing the likelihood function of a Gaussian mixture model (GMM) density ﬁ tted to the data: . p(yi |µ, Σ) = ∑ j p(yi |µ j , Σ j ) p j = ∑ j N (yi ; µ j , Σ j ) p j . (1) Each gaussian component deﬁ a local neighborhood centered around µ j with axes denes ﬁ ned by the eigenvectors of Σ j . The amount of data variance along each axis is indicated by the eigenvalues of Σ j ; if the data manifold is locally linear in the vicinity of the µ j , all but the d dominant eigenvalues will be near-zero, implying that the associated eigenvectors constitute the optimal variance-preserving local coordinate system. To some degree likelihood maximization will naturally realize this property: It requires that the GMM components shrink in volume to ﬁ the data as tightly as possible, which is best achieved by t positioning the components so that they “ pancake” onto locally ﬂat collections of datapoints. However, this state of affairs is easily violated by degenerate (zero-variance) GMM components or components ﬁ tted to overly small enough locales where the data density off the manifold is comparable to density on the manifold (e.g., at the noise scale). Consequently a prior is needed. Criterion (2) implies that neighboring partitions should have dominant axes that span similar subspaces, since disagreement (large subspace angles) would lead to inconsistent projections of a point and therefore uncertainty about its location in a low-dimensional coordinate space. The principal insight is that criterion (2) is exactly the cost of coding the location of a point in one neighborhood when it is generated by another neighborhood— the cross-entropy between the gaussian models deﬁ ning the two neighborhoods: D(N1 N2 ) = = dy N (y; µ1 ,Σ1 ) log N (y; µ1 ,Σ1 ) N (y; µ2 ,Σ2 ) (log |Σ−1 Σ2 | + trace(Σ−1 Σ1 ) + (µ2 −µ1 ) Σ−1 (µ2 −µ1 ) − D)/2. (2) 1 2 2 Roughly speaking, the terms in (2) measure differences in size, orientation, and position, respectively, of two coordinate frames located at the means µ1 , µ2 with axes speciﬁ by ed the eigenvectors of Σ1 , Σ2 . All three terms decline to zero as the overlap between the two frames is maximized. To maximize consistency between adjacent neighborhoods, we form . the prior p(µ, Σ) = exp[− ∑i= j mi (µ j )D(Ni N j )], where mi (µ j ) is a measure of co-locality. Unlike global coordination [8], we are not asking that the dominant axes in neighboring charts are aligned— only that they span nearly the same subspace. This is a much easier objective to satisfy, and it contains a useful special case where the posterior p(µ, Σ|Y) ∝ ∑i p(yi |µ, Σ)p(µ, Σ) is unimodal and can be maximized in closed form: Let us associate a gaussian neighborhood with each data-point, setting µi = yi ; take all neighborhoods to be a priori equally probable, setting pi = 1/N; and let the co-locality measure be determined from some local kernel. For example, in this paper we use mi (µ j ) ∝ N (µ j ; µi , σ2 ), with the scale parameter σ specifying the expected size of a neighborhood on the manifold in sample space. A reasonable choice is σ = r/2, so that 2erf(2) > 99.5% of the density of mi (µ j ) is contained in the area around yi where the manifold is expected to be locally linear. With uniform pi and µi , mi (µ j ) and ﬁ xed, the MAP estimates of the GMM covariances are Σi = ∑ mi (µ j ) (y j − µi )(y j − µi ) + (µ j − µi )(µ j − µi ) + Σ j j ∑ mi (µ j ) (3) . j Note that each covariance Σi is dependent on all other Σ j . The MAP estimators for all covariances can be arranged into a set of fully constrained linear equations and solved exactly for their mutually optimal values. This key step brings nonlocal information about the manifold’s shape into the local description of each neighborhood, ensuring that adjoining neighborhoods have similar covariances and small angles between their respective subspaces. Even if a local subset of data points are dense in a direction perpendicular to the manifold, the prior encourages the local chart to orient parallel to the manifold as part of a globally optimal solution, protecting against a pathology noted in [8]. Equation (3) is easily adapted to give a reduced number of charts and/or charts centered on local centroids. 4 Connecting the charts We now build a connection for set of charts speciﬁ as an arbitrary nondegenerate GMM. A ed GMM gives a soft partitioning of the dataset into neighborhoods of mean µk and covariance Σk . The optimal variance-preserving low-dimensional coordinate system for each neighborhood derives from its weighted principal component analysis, which is exactly speciﬁ ed by the eigenvectors of its covariance matrix: Eigendecompose Vk Λk Vk ← Σk with eigen. values in descending order on the diagonal of Λk and let Wk = [Id , 0]Vk be the operator . th projecting points into the k local chart, such that local chart coordinate uki = Wk (yi − µk ) . and Uk = [uk1 , · · · , ukN ] holds the local coordinates of all points. Our goal is to sew together all charts into a globally consistent low-dimensional coordinate system. For each chart there will be a low-dimensional afﬁ transform Gk ∈ R(d+1)×d ne that projects Uk into the global coordinate space. Summing over all charts, the weighted average of the projections of point yi into the low-dimensional vector space is W j (y − µ j ) 1 . x|y = ∑ G j j p j|y (y) . xi |yi = ∑ G j ⇒ u ji 1 j p j|y (yi ), (4) where pk|y (y) ∝ pk N (y; µk , Σk ), ∑k pk|y (y) = 1 is the probability that chart k generates point y. As pointed out in [8], if a point has nonzero probabilities in two charts, then there should be afﬁ transforms of those two charts that map the point to the same place in a ne global coordinate space. We set this up as a weighted least-squares problem: . G = [G1 , · · · , GK ] = arg min uki 1 ∑ pk|y (yi )p j|y (yi ) Gk Gk ,G j i −Gj u ji 1 2 . (5) F Equation (5) generates a homogeneous set of equations that determines a solution up to an afﬁ transform of G. There are two solution methods. First, let us temporarily anchor one ne neighborhood at the origin to ﬁ this indeterminacy. This adds the constraint G1 = [I, 0] . x . To solve, deﬁ indicator matrix Fk = [0, · · · , 0, I, 0, · · · , 0] with the identity mane . trix occupying the kth block, such that Gk = GFk . Let the diagonal of Pk = diag([pk|y (y1 ), · · · , pk|y (yN )]) record the per-point posteriors of chart k. The squared error of the connection is then a sum of of all patch-to-anchor and patch-to-patch inconsistencies: . E =∑ (GUk − k U1 0 2 )Pk P1 F + ∑ (GU j − GUk )P j Pk j=k 2 F ; . Uk = Fk Uk 1 . (6) Setting dE /dG = 0 and solving to minimize convex E gives −1 G = ∑ Uk P2 k k ∑ j=k P2 j Uk − ∑ ∑ Uk P2 P2 k 1 Uk P2 P2 U j k j k j=k U1 0 . (7) We now remove the dependence on a reference neighborhood G1 by rewriting equation 5, G = arg min ∑ j=k (GU j − GUk )P j Pk G 2 F = GQ 2 F = trace(GQQ G ) , (8) . where Q = ∑ j=k U j − Uk P j Pk . If we require that GG = I to prevent degenerate solutions, then equation (8) is solved (up to rotation in coordinate space) by setting G to the eigenvectors associated with the smallest eigenvalues of QQ . The eigenvectors can be computed efﬁ ciently without explicitly forming QQ ; other numerical efﬁ ciencies obtain by zeroing any vanishingly small probabilities in each Pk , yielding a sparse eigenproblem. A more interesting strategy is to numerically condition the problem by calculating the trailing eigenvectors of QQ + 1. It can be shown that this maximizes the posterior 2 p(G|Q) ∝ p(Q|G)p(G) ∝ e− GQ F e− G1 , where the prior p(G) favors a mapping G whose unit-norm rows are also zero-mean. This maximizes variance in each row of G and thereby spreads the projected points broadly and evenly over coordinate space. The solutions for MAP charts (equation (5)) and connection (equation (8)) can be applied to any well-ﬁ tted mixture of gaussians/factors1 /PCAs density model; thus large eigenproblems can be avoided by connecting just a small number of charts that cover the data. 1 We thank reviewers for calling our attention to Teh & Roweis ([11]— in this volume), which shows how to connect a set of given local dimensionality reducers in a generalized eigenvalue problem that is related to equation (8). LLE, n=5 charting (projection onto coordinate space) charting best Isomap LLE, n=6 LLE, n=7 LLE, n=8 random subset of local charts XYZ view LLE, n=9 LLE, n=10 XZ view data (linked) embedding, XY view XY view original data reconstruction (back−projected coordinate grid) best LLE (regularized) Figure 2: The twisted curl problem. L EFT: Comparison of charting, I SO M AP, & LLE. 400 points are randomly sampled from the manifold with noise. Charting is the only method that recovers the original space without catastrophes (folding), albeit with some shear. R IGHT: The manifold is regularly sampled (with noise) to illustrate the forward and backward projections. Samples are shown linked into lines to help visualize the manifold structure. Coordinate axes of a random selection of charts are shown as bold lines. Connecting subsets of charts such as this will also give good mappings. The upper right quadrant shows various LLE results. At bottom we show the charting solution and the reconstructed (back-projected) manifold, which smooths out the noise. Once the connection is solved, equation (4) gives the forward projection of any point y down into coordinate space. There are several numerically distinct candidates for the backprojection: posterior mean, mode, or exact inverse. In general, there may not be a unique posterior mode and the exact inverse is not solvable in closed form (this is also true of [8]). Note that chart-wise projection deﬁ a complementary density in coordinate space nes px|k (x) = N (x; Gk 0 1 , Gk [Id , 0]Λk [Id , 0] 0 0 0 Gk ). (9) Let p(y|x, k), used to map x into subspace k on the surface of the manifold, be a Dirac delta function whose mean is a linear function of x. Then the posterior mean back-projection is obtained by integrating out uncertainty over which chart generates x: y|x = ∑ pk|x (x) k µk + Wk Gk I 0 + x − Gk 0 1 , (10) where (·)+ denotes pseudo-inverse. In general, a back-projecting map should not reconstruct the original points. Instead, equation (10) generates a surface that passes through the weighted average of the µi of all the neighborhoods in which yi has nonzero probability, much like a principal curve passes through the center of each local group of points. 5 Experiments Synthetic examples: 400 2 D points were randomly sampled from a 2 D square and embedded in 3 D via a curl and twist, then contaminated with gaussian noise. Even if noiselessly sampled, this manifold cannot be “ unrolled” without distortion. In addition, the outer curl is sampled much less densely than the inner curl. With an order of magnitude fewer points, higher noise levels, no possibility of an isometric mapping, and uneven sampling, this is arguably a much more challenging problem than the “ swiss roll” and “ s-curve” problems featured in [12, 9, 8, 1]. Figure 2LEFT contrasts the (unique) output of charting and the best outputs obtained from I SO M AP and LLE (considering all neighborhood sizes between 2 and 20 points). I SO M AP and LLE show catastrophic folding; we had to change LLE’s b. data, yz view c. local charts d. 2D embedding e. 1D embedding 1D ordinate a. data, xy view true manifold arc length Figure 3: Untying a trefoil knot ( ) by charting. 900 noisy samples from a 3 D-embedded 1 D manifold are shown as connected dots in front (a) and side (b) views. A subset of charts is shown in (c). Solving for the 2 D connection gives the “ unknot” in (d). After removing some points to cut the knot, charting gives a 1 D embedding which we plot against true manifold arc length in (e); monotonicity (modulo noise) indicates correctness. Three principal degrees of freedom recovered from raw jittered images pose scale expression images synthesized via backprojection of straight lines in coordinate space Figure 4: Modeling the manifold of facial images from raw video. Each row contains images synthesized by back-projecting an axis-parallel straight line in coordinate space onto the manifold in image space. Blurry images correspond to points on the manifold whose neighborhoods contain few if any nearby data points. regularization in order to coax out nondegenerate (>1 D) solutions. Although charting is not designed for isometry, after afﬁ transform the forward-projected points disagree with ne the original points with an RMS error of only 1.0429, lower than the best LLE (3.1423) or best I SO M AP (1.1424, not shown). Figure 2RIGHT shows the same problem where points are sampled regularly from a grid, with noise added before and after embedding. Figure 3 shows a similar treatment of a 1 D line that was threaded into a 3 D trefoil knot, contaminated with gaussian noise, and then “ untied” via charting. Video: We obtained a 1965-frame video sequence (courtesy S. Roweis and B. Frey) of 20 × 28-pixel images in which B.F. strikes a variety of poses and expressions. The video is heavily contaminated with synthetic camera jitters. We used raw images, though image processing could have removed this and other uninteresting sources of variation. We took a 500-frame subsequence and left-right mirrored it to obtain 1000 points in 20 × 28 = 560D image space. The point-growth process peaked just above d = 3 dimensions. We solved for 25 charts, each centered on a random point, and a 3D connection. The recovered degrees of freedom— recognizable as pose, scale, and expression— are visualized in ﬁ gure 4. original data stereographic map to 3D fishbowl charting Figure 5: Flattening a ﬁ shbowl. From the left: Original 2000×2D points; their stereographic mapping to a 3D ﬁ shbowl; its 2D embedding recovered using 500 charts; and the stereographic map. Fewer charts lead to isometric mappings that fold the bowl (not shown). Conformality: Some manifolds can be ﬂattened conformally (preserving local angles) but not isometrically. Figure 5 shows that if the data is ﬁ nely charted, the connection behaves more conformally than isometrically. This problem was suggested by J. Tenenbaum. 6 Discussion Charting breaks kernel-based NLDR into two subproblems: (1) Finding a set of datacovering locally linear neighborhoods (“ charts” ) such that adjoining neighborhoods span maximally similar subspaces, and (2) computing a minimal-distortion merger (“ connection” ) of all charts. The solution to (1) is optimal w.r.t. the estimated scale of local linearity r; the solution to (2) is optimal w.r.t. the solution to (1) and the desired dimensionality d. Both problems have Bayesian settings. By ofﬂoading the nonlinearity onto the kernels, we obtain least-squares problems and closed form solutions. This scheme is also attractive because large eigenproblems can be avoided by using a reduced set of charts. The dependence on r, like trusted-set methods, is a potential source of solution instability. In practice the point-growth estimate seems fairly robust to data perturbations (to be expected if the data density changes slowly over a manifold of integral Hausdorff dimension), while the use of a soft neighborhood partitioning appears to make charting solutions reasonably stable to variations in r. Eigenvalue stability analyses may prove useful here. Ultimately, we would prefer to integrate r out. In contrast, use of d appears to be a virtue: Unlike other eigenvector-based methods, the best d-dimensional embedding is not merely a linear projection of the best d + 1-dimensional embedding; a unique distortion is found for each value of d that maximizes the information content of its embedding. Why does charting performs well on datasets where the signal-to-noise ratio confounds recent state-of-the-art methods? Two reasons may be adduced: (1) Nonlocal information is used to construct both the system of local charts and their global connection. (2) The mapping only preserves the component of local point-to-point distances that project onto the manifold; relationships perpendicular to the manifold are discarded. Thus charting uses global shape information to suppress noise in the constraints that determine the mapping. Acknowledgments Thanks to J. Buhmann, S. Makar, S. Roweis, J. Tenenbaum, and anonymous reviewers for insightful comments and suggested “ challenge” problems. References [1] M. Balasubramanian and E. L. Schwartz. The IsoMap algorithm and topological stability. Science, 295(5552):7, January 2002. [2] C. Bregler and S. Omohundro. Nonlinear image interpolation using manifold learning. In NIPS–7, 1995. [3] D. DeMers and G. Cottrell. Nonlinear dimensionality reduction. In NIPS–5, 1993. [4] J. Gomes and A. Mojsilovic. A variational approach to recovering a manifold from sample points. In ECCV, 2002. [5] T. Hastie and W. Stuetzle. Principal curves. J. Am. Statistical Assoc, 84(406):502–516, 1989. [6] G. Hinton, P. Dayan, and M. Revow. Modeling the manifolds of handwritten digits. IEEE Trans. Neural Networks, 8, 1997. [7] N. Kambhatla and T. Leen. Dimensionality reduction by local principal component analysis. Neural Computation, 9, 1997. [8] S. Roweis, L. Saul, and G. Hinton. Global coordination of linear models. In NIPS–13, 2002. [9] S. T. Roweis and L. K. Saul. Nonlinear dimensionality reduction by locally linear embedding. Science, 290:2323–2326, December 22 2000. [10] A. Smola, S. Mika, B. Schölkopf, and R. Williamson. Regularized principal manifolds. Machine Learning, 1999. [11] Y. W. Teh and S. T. Roweis. Automatic alignment of hidden representations. In NIPS–15, 2003. [12] J. B. Tenenbaum, V. de Silva, and J. C. Langford. A global geometric framework for nonlinear dimensionality reduction. Science, 290:2319–2323, December 22 2000.

Reference: text

Summary: the most important sentenses genereted by tfidf model

sentIndex sentText sentNum sentScore

1 com/people/brand/ Abstract We construct a nonlinear mapping from a high-dimensional sample space to a low-dimensional vector space, effectively recovering a Cartesian coordinate system for the manifold from which the data is sampled. [sent-3, score-0.761]

2 The mapping preserves local geometric relations in the manifold and is pseudo-invertible. [sent-4, score-0.552]

3 1 Nonlinear dimensionality reduction (NLDR) by charting Charting is the problem of assigning a low-dimensional coordinate system to data points in a high-dimensional sample space. [sent-7, score-0.731]

4 It is presumed that the data lies on or near a lowdimensional manifold embedded in the sample space, and that there exists a 1-to-1 smooth nonlinear transform between the manifold and a low-dimensional vector space. [sent-8, score-0.985]

5 The datamodeler’s goal is to estimate smooth continuous mappings between the sample and coordinate spaces. [sent-9, score-0.29]

6 This constraint is expressed in a posterior that combines a standard gaussian mixture model (GMM) likelihood function with a prior that penalizes uncertainty due to inconsistent projections in the mixture. [sent-12, score-0.095]

7 Section 3 develops a special case where this posterior is unimodal and maximizable in closed form, yielding a GMM whose covariances reveal a patchwork of overlapping locally linear subspaces that cover the manifold. [sent-13, score-0.302]

8 Section 4 shows that for this (or any) GMM and a choice of reduced dimension d, there is a unique, closed-form solution for a minimally distorting merger of the subspaces into a d-dimensional coordinate space, as well as an reverse mapping deﬁning the surface of the manifold in the sample space. [sent-14, score-0.956]

9 The intrinsic dimensionality d of the data manifold can be estimated from the growth process of point-to-point distances. [sent-15, score-0.629]

10 In analogy to differential geometry, we call the subspaces “charts” and their merger the “connection. [sent-16, score-0.113]

11 The ﬁ has its roots in mapeld ping algorithms: DeMers and Cottrell [3] proposed using auto-encoding neural networks with a hidden layer “ bottleneck,” effectively casting dimensionality reduction as a compression problem. [sent-20, score-0.099]

12 More recent advances aim for embeddings: Gomes and Mojsilovic [4] treat manifold completion as an anisotropic diffusion problem, iteratively expanding points until they connect to their neighbors. [sent-24, score-0.505]

13 The I SO M AP algorithm [12] represents remote distances as sums of a trusted set of distances between immediate neighbors, then uses multidimensional scaling to compute a low-dimensional embedding that minimally distorts all distances. [sent-25, score-0.201]

14 The locally linear embedding algorithm (LLE) [9] represents each point as a weighted combination of a trusted set of nearest neighbors, then computes a minimally distorting low-dimensional barycentric embedding. [sent-26, score-0.355]

15 In a return to mapping, Roweis and colleagues [8] proposed global coordination— learning a mixture of locally linear projections from sample to coordinate space. [sent-30, score-0.469]

16 Like [2, 7, 6, 8], the method we develop below is a decomposition of the manifold into locally linear neighborhoods. [sent-33, score-0.536]

17 It bears closest relation to global coordination [8], although by a different construction of the problem, we avoid hill-climbing a spiky posterior and instead develop a closed-form solution. [sent-34, score-0.135]

18 2 Estimating locally linear scale and intrinsic dimensionality . [sent-35, score-0.347]

19 We begin with matrix of sample points Y = [y1 , · · · , yN ], yn ∈ RD populating a Ddimensional sample space, and a conjecture that these points are samples from a manifold M of intrinsic dimensionality d < D. [sent-36, score-0.856]

20 G(Y) → X = [x1 , · · · , xN ], xn ∈ Rd and 1-to-1 reverse mapping G−1 (X) → Y such that local relations between nearby points are preserved (this will be formalized below). [sent-38, score-0.274]

21 The map G should be non-catastrophic, that is, without folds: Parallel lines on the manifold in RD should map to continuous smooth non-intersecting curves in Rd . [sent-39, score-0.492]

22 Smoothness means that at some scale r the mapping from a neighborhood on M to Rd is effectively linear. [sent-41, score-0.265]

23 The count n(r) grows as rd , but only at the locally linear scale; the grow rate is inﬂated by isotropic noise at smaller scales and by embedding curvature at larger scales. [sent-43, score-0.418]

24 To estimate r, we look at how the r-ball grows as points are added to it, tracking c(r) = d d log n(r) log r. [sent-45, score-0.123]

25 At noise scales, c(r) ≈ 1/D < 1/d, because noise has distributed points in all directions with equal probability. [sent-46, score-0.179]

26 At the scale at which curvature becomes signiﬁ cant, c(r) < 1/d, because the manifold is no longer perpendicular to the surface of the ball, so the ball does not have to grow as fast to accommodate new points. [sent-47, score-0.661]

27 At the locally linear scale, the process peaks at c(r) = 1/d, because points are distributed only in the directions of the manifold’s local tangent space. [sent-48, score-0.277]

28 L EFT: At the locally linear scale, the number of points in an r-ball grows as rd ; at noise and curvature scales it grows faster. [sent-50, score-0.445]

29 R IGHT: Using the point-count growth process to ﬁ the intrinsic dimensionality of a 2D manifold nonlinearly nd embedded in 3-space (see ﬁ gure 2). [sent-51, score-0.629]

30 For neighborhoods of radius r ≈ 1 with roughly n ≈ 10 points, the slope peaks at 1/2 indicating a dimensionality of d = 2. [sent-53, score-0.258]

31 Below that, the data appears 3 D because it is dominated by noise (except for n ≤ D points); above, the data appears >2 D because of manifold curvature. [sent-54, score-0.464]

32 As the r-ball expands to cover the entire data-set the dimensionality appears to drop to 1 as the process begins to track the 1D edges of the 2D sheet. [sent-55, score-0.099]

33 For low-dimensional manifolds such as sheets, the boundary submanifolds (edges and corners) are very small relative to the full manifold, so the boundary effect is typically limited to a small rise in c(r) as r approaches the scale of the entire data set. [sent-57, score-0.115]

34 3 Charting the data In the charting step we ﬁ a soft partitioning of the data into locally linear low-dimensional nd neighborhoods, as a prelude to computing the connection that gives the global lowdimensional embedding. [sent-60, score-0.607]

35 To minimize information loss in the connection, we require that the data points project into a subspace associated with each neighborhood with (1) minimal loss of local variance and (2) maximal agreement of the projections of nearby points into nearby neighborhoods. [sent-61, score-0.531]

36 (1) Each gaussian component deﬁ a local neighborhood centered around µ j with axes denes ﬁ ned by the eigenvectors of Σ j . [sent-64, score-0.343]

37 To some degree likelihood maximization will naturally realize this property: It requires that the GMM components shrink in volume to ﬁ the data as tightly as possible, which is best achieved by t positioning the components so that they “ pancake” onto locally ﬂat collections of datapoints. [sent-66, score-0.153]

38 However, this state of affairs is easily violated by degenerate (zero-variance) GMM components or components ﬁ tted to overly small enough locales where the data density off the manifold is comparable to density on the manifold (e. [sent-67, score-0.956]

39 Criterion (2) implies that neighboring partitions should have dominant axes that span similar subspaces, since disagreement (large subspace angles) would lead to inconsistent projections of a point and therefore uncertainty about its location in a low-dimensional coordinate space. [sent-71, score-0.36]

40 (2) 1 2 2 Roughly speaking, the terms in (2) measure differences in size, orientation, and position, respectively, of two coordinate frames located at the means µ1 , µ2 with axes speciﬁ by ed the eigenvectors of Σ1 , Σ2 . [sent-73, score-0.321]

41 the prior p(µ, Σ) = exp[− ∑i= j mi (µ j )D(Ni N j )], where mi (µ j ) is a measure of co-locality. [sent-76, score-0.15]

42 Unlike global coordination [8], we are not asking that the dominant axes in neighboring charts are aligned— only that they span nearly the same subspace. [sent-77, score-0.588]

43 For example, in this paper we use mi (µ j ) ∝ N (µ j ; µi , σ2 ), with the scale parameter σ specifying the expected size of a neighborhood on the manifold in sample space. [sent-79, score-0.736]

44 5% of the density of mi (µ j ) is contained in the area around yi where the manifold is expected to be locally linear. [sent-81, score-0.744]

45 With uniform pi and µi , mi (µ j ) and ﬁ xed, the MAP estimates of the GMM covariances are Σi = ∑ mi (µ j ) (y j − µi )(y j − µi ) + (µ j − µi )(µ j − µi ) + Σ j j ∑ mi (µ j ) (3) . [sent-82, score-0.264]

46 This key step brings nonlocal information about the manifold’s shape into the local description of each neighborhood, ensuring that adjoining neighborhoods have similar covariances and small angles between their respective subspaces. [sent-85, score-0.384]

47 Even if a local subset of data points are dense in a direction perpendicular to the manifold, the prior encourages the local chart to orient parallel to the manifold as part of a globally optimal solution, protecting against a pathology noted in [8]. [sent-86, score-0.834]

48 Equation (3) is easily adapted to give a reduced number of charts and/or charts centered on local centroids. [sent-87, score-0.899]

49 4 Connecting the charts We now build a connection for set of charts speciﬁ as an arbitrary nondegenerate GMM. [sent-88, score-0.904]

50 A ed GMM gives a soft partitioning of the dataset into neighborhoods of mean µk and covariance Σk . [sent-89, score-0.191]

51 The optimal variance-preserving low-dimensional coordinate system for each neighborhood derives from its weighted principal component analysis, which is exactly speciﬁ ed by the eigenvectors of its covariance matrix: Eigendecompose Vk Λk Vk ← Σk with eigen. [sent-90, score-0.451]

52 th projecting points into the k local chart, such that local chart coordinate uki = Wk (yi − µk ) . [sent-92, score-0.596]

53 Our goal is to sew together all charts into a globally consistent low-dimensional coordinate system. [sent-94, score-0.638]

54 For each chart there will be a low-dimensional afﬁ transform Gk ∈ R(d+1)×d ne that projects Uk into the global coordinate space. [sent-95, score-0.407]

55 Summing over all charts, the weighted average of the projections of point yi into the low-dimensional vector space is W j (y − µ j ) 1 . [sent-96, score-0.147]

56 xi |yi = ∑ G j ⇒ u ji 1 j p j|y (yi ), (4) where pk|y (y) ∝ pk N (y; µk , Σk ), ∑k pk|y (y) = 1 is the probability that chart k generates point y. [sent-98, score-0.325]

57 As pointed out in [8], if a point has nonzero probabilities in two charts, then there should be afﬁ transforms of those two charts that map the point to the same place in a ne global coordinate space. [sent-99, score-0.692]

58 First, let us temporarily anchor one ne neighborhood at the origin to ﬁ this indeterminacy. [sent-104, score-0.126]

59 Let the diagonal of Pk = diag([pk|y (y1 ), · · · , pk|y (yN )]) record the per-point posteriors of chart k. [sent-109, score-0.115]

60 (7) We now remove the dependence on a reference neighborhood G1 by rewriting equation 5, G = arg min ∑ j=k (GU j − GUk )P j Pk G 2 F = GQ 2 F = trace(GQQ G ) , (8) . [sent-114, score-0.126]

61 If we require that GG = I to prevent degenerate solutions, then equation (8) is solved (up to rotation in coordinate space) by setting G to the eigenvectors associated with the smallest eigenvalues of QQ . [sent-116, score-0.269]

62 It can be shown that this maximizes the posterior 2 p(G|Q) ∝ p(Q|G)p(G) ∝ e− GQ F e− G1 , where the prior p(G) favors a mapping G whose unit-norm rows are also zero-mean. [sent-119, score-0.104]

63 This maximizes variance in each row of G and thereby spreads the projected points broadly and evenly over coordinate space. [sent-120, score-0.296]

64 The solutions for MAP charts (equation (5)) and connection (equation (8)) can be applied to any well-ﬁ tted mixture of gaussians/factors1 /PCAs density model; thus large eigenproblems can be avoided by connecting just a small number of charts that cover the data. [sent-121, score-1.015]

65 1 We thank reviewers for calling our attention to Teh & Roweis ([11]— in this volume), which shows how to connect a set of given local dimensionality reducers in a generalized eigenvalue problem that is related to equation (8). [sent-122, score-0.171]

66 400 points are randomly sampled from the manifold with noise. [sent-125, score-0.542]

67 R IGHT: The manifold is regularly sampled (with noise) to illustrate the forward and backward projections. [sent-127, score-0.455]

68 Samples are shown linked into lines to help visualize the manifold structure. [sent-128, score-0.454]

69 Coordinate axes of a random selection of charts are shown as bold lines. [sent-129, score-0.449]

70 Connecting subsets of charts such as this will also give good mappings. [sent-130, score-0.397]

71 At bottom we show the charting solution and the reconstructed (back-projected) manifold, which smooths out the noise. [sent-132, score-0.296]

72 Once the connection is solved, equation (4) gives the forward projection of any point y down into coordinate space. [sent-133, score-0.278]

73 Note that chart-wise projection deﬁ a complementary density in coordinate space nes px|k (x) = N (x; Gk 0 1 , Gk [Id , 0]Λk [Id , 0] 0 0 0 Gk ). [sent-136, score-0.248]

74 Then the posterior mean back-projection is obtained by integrating out uncertainty over which chart generates x: y|x = ∑ pk|x (x) k µk + Wk Gk I 0 + x − Gk 0 1 , (10) where (·)+ denotes pseudo-inverse. [sent-138, score-0.188]

75 Instead, equation (10) generates a surface that passes through the weighted average of the µi of all the neighborhoods in which yi has nonzero probability, much like a principal curve passes through the center of each local group of points. [sent-140, score-0.447]

76 5 Experiments Synthetic examples: 400 2 D points were randomly sampled from a 2 D square and embedded in 3 D via a curl and twist, then contaminated with gaussian noise. [sent-141, score-0.226]

77 Even if noiselessly sampled, this manifold cannot be “ unrolled” without distortion. [sent-142, score-0.418]

78 In addition, the outer curl is sampled much less densely than the inner curl. [sent-143, score-0.099]

79 Figure 2LEFT contrasts the (unique) output of charting and the best outputs obtained from I SO M AP and LLE (considering all neighborhood sizes between 2 and 20 points). [sent-145, score-0.422]

80 data, xy view true manifold arc length Figure 3: Untying a trefoil knot ( ) by charting. [sent-151, score-0.575]

81 900 noisy samples from a 3 D-embedded 1 D manifold are shown as connected dots in front (a) and side (b) views. [sent-152, score-0.418]

82 After removing some points to cut the knot, charting gives a 1 D embedding which we plot against true manifold arc length in (e); monotonicity (modulo noise) indicates correctness. [sent-155, score-0.93]

83 Three principal degrees of freedom recovered from raw jittered images pose scale expression images synthesized via backprojection of straight lines in coordinate space Figure 4: Modeling the manifold of facial images from raw video. [sent-156, score-0.909]

84 Each row contains images synthesized by back-projecting an axis-parallel straight line in coordinate space onto the manifold in image space. [sent-157, score-0.698]

85 Blurry images correspond to points on the manifold whose neighborhoods contain few if any nearby data points. [sent-158, score-0.753]

86 Although charting is not designed for isometry, after afﬁ transform the forward-projected points disagree with ne the original points with an RMS error of only 1. [sent-160, score-0.504]

87 Figure 2RIGHT shows the same problem where points are sampled regularly from a grid, with noise added before and after embedding. [sent-164, score-0.17]

88 original data stereographic map to 3D fishbowl charting Figure 5: Flattening a ﬁ shbowl. [sent-177, score-0.387]

89 From the left: Original 2000×2D points; their stereographic mapping to a 3D ﬁ shbowl; its 2D embedding recovered using 500 charts; and the stereographic map. [sent-178, score-0.266]

90 Fewer charts lead to isometric mappings that fold the bowl (not shown). [sent-179, score-0.474]

91 Conformality: Some manifolds can be ﬂattened conformally (preserving local angles) but not isometrically. [sent-180, score-0.146]

92 Figure 5 shows that if the data is ﬁ nely charted, the connection behaves more conformally than isometrically. [sent-181, score-0.105]

93 6 Discussion Charting breaks kernel-based NLDR into two subproblems: (1) Finding a set of datacovering locally linear neighborhoods (“ charts” ) such that adjoining neighborhoods span maximally similar subspaces, and (2) computing a minimal-distortion merger (“ connection” ) of all charts. [sent-184, score-0.531]

94 the estimated scale of local linearity r; the solution to (2) is optimal w. [sent-188, score-0.149]

95 the solution to (1) and the desired dimensionality d. [sent-191, score-0.099]

96 Why does charting performs well on datasets where the signal-to-noise ratio confounds recent state-of-the-art methods? [sent-200, score-0.296]

97 Two reasons may be adduced: (1) Nonlocal information is used to construct both the system of local charts and their global connection. [sent-201, score-0.518]

98 (2) The mapping only preserves the component of local point-to-point distances that project onto the manifold; relationships perpendicular to the manifold are discarded. [sent-202, score-0.625]

99 Thus charting uses global shape information to suppress noise in the constraints that determine the mapping. [sent-203, score-0.391]

100 A variational approach to recovering a manifold from sample points. [sent-228, score-0.458]

similar papers computed by tfidf model

tfidf for this paper:

wordName wordTfidf (topN-words)

[('manifold', 0.418), ('charts', 0.397), ('charting', 0.296), ('coordinate', 0.209), ('lle', 0.184), ('gk', 0.18), ('pk', 0.179), ('gmm', 0.166), ('neighborhoods', 0.159), ('neighborhood', 0.126), ('locally', 0.118), ('chart', 0.115), ('dimensionality', 0.099), ('embedding', 0.096), ('yi', 0.094), ('points', 0.087), ('scale', 0.077), ('mi', 0.075), ('rd', 0.073), ('uk', 0.073), ('local', 0.072), ('nldr', 0.072), ('ap', 0.071), ('connection', 0.069), ('mapping', 0.062), ('curl', 0.062), ('guk', 0.062), ('trusted', 0.062), ('eigenvectors', 0.06), ('roweis', 0.06), ('subspaces', 0.059), ('growth', 0.059), ('principal', 0.056), ('merger', 0.054), ('knot', 0.054), ('stereographic', 0.054), ('intrinsic', 0.053), ('nearby', 0.053), ('projections', 0.053), ('axes', 0.052), ('wk', 0.049), ('curvature', 0.049), ('global', 0.049), ('dominant', 0.046), ('qq', 0.046), ('vk', 0.046), ('noise', 0.046), ('closed', 0.044), ('coordination', 0.044), ('video', 0.044), ('ball', 0.044), ('minimally', 0.043), ('lowdimensional', 0.043), ('tted', 0.042), ('posterior', 0.042), ('adjoining', 0.041), ('backprojection', 0.041), ('eft', 0.041), ('eigenproblems', 0.041), ('gomes', 0.041), ('nondegenerate', 0.041), ('nonlocal', 0.041), ('uki', 0.041), ('mappings', 0.041), ('sample', 0.04), ('contaminated', 0.04), ('density', 0.039), ('af', 0.039), ('covariances', 0.039), ('manifolds', 0.038), ('perpendicular', 0.038), ('map', 0.037), ('sampled', 0.037), ('grows', 0.036), ('trefoil', 0.036), ('conformally', 0.036), ('demers', 0.036), ('distorting', 0.036), ('folding', 0.036), ('ight', 0.036), ('isometric', 0.036), ('sheets', 0.036), ('visualize', 0.036), ('images', 0.036), ('id', 0.035), ('surface', 0.035), ('onto', 0.035), ('xy', 0.034), ('transform', 0.034), ('centered', 0.033), ('arc', 0.033), ('gu', 0.033), ('yn', 0.032), ('angles', 0.032), ('partitioning', 0.032), ('globally', 0.032), ('nonlinear', 0.032), ('gq', 0.031), ('generates', 0.031), ('connecting', 0.03)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 1.0000006 49 nips-2002-Charting a Manifold

Author: Matthew Brand

2 0.26622611 36 nips-2002-Automatic Alignment of Local Representations

Author: Yee W. Teh, Sam T. Roweis

Abstract: We present an automatic alignment procedure which maps the disparate internal representations learned by several local dimensionality reduction experts into a single, coherent global coordinate system for the original data space. Our algorithm can be applied to any set of experts, each of which produces a low-dimensional local representation of a highdimensional input. Unlike recent efforts to coordinate such models by modifying their objective functions [1, 2], our algorithm is invoked after training and applies an efﬁcient eigensolver to post-process the trained models. The post-processing has no local optima and the size of the system it must solve scales with the number of local models rather than the number of original data points, making it more efﬁcient than model-free algorithms such as Isomap [3] or LLE [4]. 1 Introduction: Local vs. Global Dimensionality Reduction Beyond density modelling, an important goal of unsupervised learning is to discover compact, informative representations of high-dimensional data. If the data lie on a smooth low dimensional manifold, then an excellent encoding is the coordinates internal to that manifold. The process of determining such coordinates is dimensionality reduction. Linear dimensionality reduction methods such as principal component analysis and factor analysis are easy to train but cannot capture the structure of curved manifolds. Mixtures of these simple unsupervised models [5, 6, 7, 8] have been used to perform local dimensionality reduction, and can provide good density models for curved manifolds, but unfortunately such mixtures cannot do dimensionality reduction. They do not describe a single, coherent low-dimensional coordinate system for the data since there is no pressure for the local coordinates of each component to agree. Roweis et al [1] recently proposed a model which performs global coordination of local coordinate systems in a mixture of factor analyzers (MFA). Their model is trained by maximizing the likelihood of the data, with an additional variational penalty term to encourage the internal coordinates of the factor analyzers to agree. While their model can trade off modelling the data and having consistent local coordinate systems, it requires a user given trade-off parameter, training is quite inefﬁcient (although [2] describes an improved training algorithm for a more constrained model), and it has quite serious local minima problems (methods like LLE [4] or Isomap [3] have to be used for initialization). In this paper we describe a novel, automatic way to align the hidden representations used by each component of a mixture of dimensionality reducers into a single global representation of the data throughout space. Given an already trained mixture, the alignment is achieved by applying an eigensolver to a matrix constructed from the internal representations of the mixture components. Our method is efﬁcient, simple to implement, and has no local optima in its optimization nor any learning rates or annealing schedules. 2 The Locally Linear Coordination Algorithm H 9¥ EI¡ CD66B9 ©9B 766 % G F 5 #

3 0.21475179 97 nips-2002-Global Versus Local Methods in Nonlinear Dimensionality Reduction

Author: Vin D. Silva, Joshua B. Tenenbaum

Abstract: Recently proposed algorithms for nonlinear dimensionality reduction fall broadly into two categories which have different advantages and disadvantages: global (Isomap [1]), and local (Locally Linear Embedding [2], Laplacian Eigenmaps [3]). We present two variants of Isomap which combine the advantages of the global approach with what have previously been exclusive advantages of local methods: computational sparsity and the ability to invert conformal maps.

4 0.18288222 138 nips-2002-Manifold Parzen Windows

Author: Pascal Vincent, Yoshua Bengio

Abstract: The similarity between objects is a fundamental element of many learning algorithms. Most non-parametric methods take this similarity to be ﬁxed, but much recent work has shown the advantages of learning it, in particular to exploit the local invariances in the data or to capture the possibly non-linear manifold on which most of the data lies. We propose a new non-parametric kernel density estimation method which captures the local structure of an underlying manifold through the leading eigenvectors of regularized local covariance matrices. Experiments in density estimation show signiﬁcant improvements with respect to Parzen density estimators. The density estimators can also be used within Bayes classiﬁers, yielding classiﬁcation rates similar to SVMs and much superior to the Parzen classiﬁer.

5 0.088985309 113 nips-2002-Information Diffusion Kernels

Author: Guy Lebanon, John D. Lafferty

Abstract: A new family of kernels for statistical learning is introduced that exploits the geometric structure of statistical models. Based on the heat equation on the Riemannian manifold deﬁned by the Fisher information metric, information diffusion kernels generalize the Gaussian kernel of Euclidean space, and provide a natural way of combining generative statistical modeling with non-parametric discriminative learning. As a special case, the kernels give a new approach to applying kernel-based learning algorithms to discrete data. Bounds on covering numbers for the new kernels are proved using spectral theory in differential geometry, and experimental results are presented for text classiﬁcation.

6 0.078419499 190 nips-2002-Stochastic Neighbor Embedding

7 0.077226102 87 nips-2002-Fast Transformation-Invariant Factor Analysis

8 0.07605163 9 nips-2002-A Minimal Intervention Principle for Coordinated Movement

9 0.070717089 19 nips-2002-Adapting Codes and Embeddings for Polychotomies

10 0.066099361 133 nips-2002-Learning to Perceive Transparency from the Statistics of Natural Scenes

11 0.065791443 117 nips-2002-Intrinsic Dimension Estimation Using Packing Numbers

12 0.06510824 98 nips-2002-Going Metric: Denoising Pairwise Data

13 0.064711034 34 nips-2002-Artefactual Structure from Least-Squares Multidimensional Scaling

14 0.063117236 82 nips-2002-Exponential Family PCA for Belief Compression in POMDPs

15 0.060147382 115 nips-2002-Informed Projections

16 0.059074987 86 nips-2002-Fast Sparse Gaussian Process Methods: The Informative Vector Machine

17 0.05698783 141 nips-2002-Maximally Informative Dimensions: Analyzing Neural Responses to Natural Signals

18 0.055552661 70 nips-2002-Distance Metric Learning with Application to Clustering with Side-Information

19 0.054750334 65 nips-2002-Derivative Observations in Gaussian Process Models of Dynamic Systems

20 0.054203317 132 nips-2002-Learning to Detect Natural Image Boundaries Using Brightness and Texture

similar papers computed by lsi model

lsi for this paper:

topicId topicWeight

[(0, -0.19), (1, -0.045), (2, -0.025), (3, 0.08), (4, -0.068), (5, 0.044), (6, -0.007), (7, 0.005), (8, -0.068), (9, 0.258), (10, 0.092), (11, 0.007), (12, 0.035), (13, -0.159), (14, 0.081), (15, 0.305), (16, -0.228), (17, 0.205), (18, 0.066), (19, -0.116), (20, -0.026), (21, 0.038), (22, -0.003), (23, -0.155), (24, -0.007), (25, 0.032), (26, -0.037), (27, -0.045), (28, -0.059), (29, -0.11), (30, 0.027), (31, -0.023), (32, -0.032), (33, -0.043), (34, 0.031), (35, 0.006), (36, 0.019), (37, -0.081), (38, -0.056), (39, -0.042), (40, 0.056), (41, -0.014), (42, -0.113), (43, -0.02), (44, -0.082), (45, -0.034), (46, -0.028), (47, -0.035), (48, 0.029), (49, -0.046)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.96986657 49 nips-2002-Charting a Manifold

Author: Matthew Brand

2 0.86462742 36 nips-2002-Automatic Alignment of Local Representations

Author: Yee W. Teh, Sam T. Roweis

3 0.86045325 97 nips-2002-Global Versus Local Methods in Nonlinear Dimensionality Reduction

Author: Vin D. Silva, Joshua B. Tenenbaum

4 0.71109074 138 nips-2002-Manifold Parzen Windows

Author: Pascal Vincent, Yoshua Bengio

5 0.64683414 117 nips-2002-Intrinsic Dimension Estimation Using Packing Numbers

Author: Balázs Kégl

Abstract: We propose a new algorithm to estimate the intrinsic dimension of data sets. The method is based on geometric properties of the data and requires neither parametric assumptions on the data generating model nor input parameters to set. The method is compared to a similar, widelyused algorithm from the same family of geometric techniques. Experiments show that our method is more robust in terms of the data generating distribution and more reliable in the presence of noise. 1

6 0.52416319 190 nips-2002-Stochastic Neighbor Embedding

7 0.40425226 34 nips-2002-Artefactual Structure from Least-Squares Multidimensional Scaling

8 0.3862583 63 nips-2002-Critical Lines in Symmetry of Mixture Models and its Application to Component Splitting

9 0.36997193 87 nips-2002-Fast Transformation-Invariant Factor Analysis

10 0.31810927 113 nips-2002-Information Diffusion Kernels

11 0.28577474 65 nips-2002-Derivative Observations in Gaussian Process Models of Dynamic Systems

12 0.28372979 115 nips-2002-Informed Projections

13 0.27395436 98 nips-2002-Going Metric: Denoising Pairwise Data

14 0.27297485 9 nips-2002-A Minimal Intervention Principle for Coordinated Movement

15 0.27173111 70 nips-2002-Distance Metric Learning with Application to Clustering with Side-Information

16 0.25260457 114 nips-2002-Information Regularization with Partially Labeled Data

17 0.23688823 18 nips-2002-Adaptation and Unsupervised Learning

18 0.23404185 47 nips-2002-Branching Law for Axons

19 0.22719131 133 nips-2002-Learning to Perceive Transparency from the Statistics of Natural Scenes

20 0.22311991 132 nips-2002-Learning to Detect Natural Image Boundaries Using Brightness and Texture

similar papers computed by lda model

lda for this paper:

topicId topicWeight

[(3, 0.013), (11, 0.015), (23, 0.013), (42, 0.078), (54, 0.22), (55, 0.024), (67, 0.015), (68, 0.361), (74, 0.068), (92, 0.036), (98, 0.063)]

similar papers list:

simIndex simValue paperId paperTitle

1 0.96588373 171 nips-2002-Reconstructing Stimulus-Driven Neural Networks from Spike Times

Author: Duane Q. Nykamp

Abstract: We present a method to distinguish direct connections between two neurons from common input originating from other, unmeasured neurons. The distinction is computed from the spike times of the two neurons in response to a white noise stimulus. Although the method is based on a highly idealized linear-nonlinear approximation of neural response, we demonstrate via simulation that the approach can work with a more realistic, integrate-and-ﬁre neuron model. We propose that the approach exempliﬁed by this analysis may yield viable tools for reconstructing stimulus-driven neural networks from data gathered in neurophysiology experiments.

same-paper 2 0.90375918 49 nips-2002-Charting a Manifold

Author: Matthew Brand

3 0.89041597 62 nips-2002-Coulomb Classifiers: Generalizing Support Vector Machines via an Analogy to Electrostatic Systems

Author: Sepp Hochreiter, Michael C. Mozer, Klaus Obermayer

Abstract: We introduce a family of classiﬁers based on a physical analogy to an electrostatic system of charged conductors. The family, called Coulomb classiﬁers, includes the two best-known support-vector machines (SVMs), the ν–SVM and the C–SVM. In the electrostatics analogy, a training example corresponds to a charged conductor at a given location in space, the classiﬁcation function corresponds to the electrostatic potential function, and the training objective function corresponds to the Coulomb energy. The electrostatic framework provides not only a novel interpretation of existing algorithms and their interrelationships, but it suggests a variety of new methods for SVMs including kernels that bridge the gap between polynomial and radial-basis functions, objective functions that do not require positive-deﬁnite kernels, regularization techniques that allow for the construction of an optimal classiﬁer in Minkowski space. Based on the framework, we propose novel SVMs and perform simulation studies to show that they are comparable or superior to standard SVMs. The experiments include classiﬁcation tasks on data which are represented in terms of their pairwise proximities, where a Coulomb Classiﬁer outperformed standard SVMs. 1

4 0.85944593 73 nips-2002-Dynamic Bayesian Networks with Deterministic Latent Tables

Author: David Barber

Abstract: The application of latent/hidden variable Dynamic Bayesian Networks is constrained by the complexity of marginalising over latent variables. For this reason either small latent dimensions or Gaussian latent conditional tables linearly dependent on past states are typically considered in order that inference is tractable. We suggest an alternative approach in which the latent variables are modelled using deterministic conditional probability tables. This specialisation has the advantage of tractable inference even for highly complex non-linear/non-Gaussian visible conditional probability tables. This approach enables the consideration of highly complex latent dynamics whilst retaining the beneﬁts of a tractable probabilistic model. 1

5 0.83101875 76 nips-2002-Dynamical Constraints on Computing with Spike Timing in the Cortex

Author: Arunava Banerjee, Alexandre Pouget

Abstract: If the cortex uses spike timing to compute, the timing of the spikes must be robust to perturbations. Based on a recent framework that provides a simple criterion to determine whether a spike sequence produced by a generic network is sensitive to initial conditions, and numerical simulations of a variety of network architectures, we argue within the limits set by our model of the neuron, that it is unlikely that precise sequences of spike timings are used for computation under conditions typically found in the cortex.

6 0.7162891 5 nips-2002-A Digital Antennal Lobe for Pattern Equalization: Analysis and Design

7 0.69924974 51 nips-2002-Classifying Patterns of Visual Motion - a Neuromorphic Approach

8 0.69701451 11 nips-2002-A Model for Real-Time Computation in Generic Neural Microcircuits

9 0.67069787 128 nips-2002-Learning a Forward Model of a Reflex

10 0.66548055 164 nips-2002-Prediction of Protein Topologies Using Generalized IOHMMs and RNNs

11 0.65660793 173 nips-2002-Recovering Intrinsic Images from a Single Image

12 0.64735401 123 nips-2002-Learning Attractor Landscapes for Learning Motor Primitives

13 0.64644575 102 nips-2002-Hidden Markov Model of Cortical Synaptic Plasticity: Derivation of the Learning Rule

14 0.63905001 137 nips-2002-Location Estimation with a Differential Update Network

15 0.63537365 4 nips-2002-A Differential Semantics for Jointree Algorithms

16 0.63460517 141 nips-2002-Maximally Informative Dimensions: Analyzing Neural Responses to Natural Signals

17 0.6336329 60 nips-2002-Convergence Properties of Some Spike-Triggered Analysis Techniques

18 0.62234724 117 nips-2002-Intrinsic Dimension Estimation Using Packing Numbers

19 0.619524 190 nips-2002-Stochastic Neighbor Embedding

20 0.61687672 186 nips-2002-Spike Timing-Dependent Plasticity in the Address Domain