Richard Nock

Statistical region merging

This paper explores a statistical basis for a process often described in computer vision: image segmentation by region merging following a particular order in the choice of regions. We exhibit a particular blend of algorithmics and statistics whose segmentation error is, as we show, limited from both the qualitative and quantitative standpoints. This approach can be efficiently approximated in linear time/space, leading to a fast segmentation algorithm tailored to processing images described using most common numerical pixel attribute spaces. The conceptual simplicity of the approach makes it simple to modify and cope with hard noise corruption, handle occlusion, authorize the control of the segmentation scale, and process unconventional data such as spherical images. Experiments on gray-level and color images, obtained with a short readily available C-code, display the quality of the segmentations obtained.

On weighting clustering

Recent papers and patents in iterative unsupervised learning have emphasized a new trend in clustering. It basically consists of penalizing solutions via weights on the instance points, somehow making clustering move toward the hardest points to cluster. The motivations come principally from an analogy with powerful supervised classification methods known as boosting algorithms. However, interest in this analogy has so far been mainly borne out from experimental studies only. This paper is, to the best of our knowledge, the first attempt at its formalization. More precisely, we handle clustering as a constrained minimization of a Bregman divergence. Weight modifications rely on the local variations of the expected complete log-likelihoods. Theoretical results show benefits resembling those of boosting algorithms and bring modified (weighted) versions of clustering algorithms such as k-means, fuzzy c-means, expectation maximization (EM), and k-harmonic means. Experiments are provided for all these algorithms, with a readily available code. They display the advantages that subtle data reweighting may bring to clustering

A hybrid filter/wrapper approach of feature selection using information theory

Abstract We focus on a hybrid approach of feature selection. We begin our analysis with a filter model , exploiting the geometrical information contained in the minimum spanning tree (MST) built on the learning set. This model exploits a statistical test of relative certainty gain , used in a forward selection algorithm. In the second part of the paper, we show that the MST can be replaced by the 1 nearest-neighbor graph without challenging the statistical framework. This leads to a feature selection algorithm belonging to a new category of hybrid models ( filter-wrapper ). Experimental results on readily available synthetic and natural domains are presented and discussed.

Sided and Symmetrized Bregman Centroids

In this paper, we generalize the notions of centroids (and barycenters) to the broad class of information-theoretic distortion measures called Bregman divergences. Bregman divergences form a rich and versatile family of distances that unifies quadratic Euclidean distances with various well-known statistical entropic measures. Since besides the squared Euclidean distance, Bregman divergences are asymmetric, we consider the left-sided and right-sided centroids and the symmetrized centroids as minimizers of average Bregman distortions. We prove that all three centroids are unique and give closed-form solutions for the sided centroids that are generalized means. Furthermore, we design a provably fast and efficient arbitrary close approximation algorithm for the symmetrized centroid based on its exact geometric characterization. The geometric approximation algorithm requires only to walk on a geodesic linking the two left/right-sided centroids. We report on our implementation for computing entropic centers of image histogram clusters and entropic centers of multivariate normal distributions that are useful operations for processing multimedia information and retrieval. These experiments illustrate that our generic methods compare favorably with former limited ad hoc methods.

Bregman Voronoi Diagrams

The Voronoi diagram of a finite set of objects is a fundamental geometric structure that subdivides the embedding space into regions, each region consisting of the points that are closer to a given object than to the others. We may define various variants of Voronoi diagrams depending on the class of objects, the distance function and the embedding space. In this paper, we investigate a framework for defining and building Voronoi diagrams for a broad class of distance functions called Bregman divergences. Bregman divergences include not only the traditional (squared) Euclidean distance but also various divergence measures based on entropic functions. Accordingly, Bregman Voronoi diagrams allow one to define information-theoretic Voronoi diagrams in statistical parametric spaces based on the relative entropy of distributions. We define several types of Bregman diagrams, establish correspondences between those diagrams (using the Legendre transformation), and show how to compute them efficiently. We also introduce extensions of these diagrams, e.g., k-order and k-bag Bregman Voronoi diagrams, and introduce Bregman triangulations of a set of points and their connection with Bregman Voronoi diagrams. We show that these triangulations capture many of the properties of the celebrated Delaunay triangulation.

Learning and evaluation in the presence of class hierarchies: application to text categorization

This paper deals with categorization tasks where categories are partially ordered to form a hierarchy. First, it introduces the notion of consistent classification which takes into account the semantics of a class hierarchy. Then, it presents a novel global hierarchical approach that produces consistent classification. This algorithm with AdaBoost as the underlying learning procedure significantly outperforms the corresponding “flat” approach, i.e. the approach that does not take into account the hierarchical information. In addition, the proposed algorithm surpasses the hierarchical local top-down approach on many synthetic and real tasks. For evaluation purposes, we use a novel hierarchical evaluation measure that has some attractive properties: it is simple, requires no parameter tuning, gives credit to partially correct classification and discriminates errors by both distance and depth in a class hierarchy.

Bregman Divergences and Surrogates for Learning

Bartlett et al. (2006) recently proved that a ground condition for surrogates, classification calibration, ties up their consistent minimization to that of the classification risk, and left as an important problem the algorithmic questions about their minimization. In this paper, we address this problem for a wide set which lies at the intersection of classification calibrated surrogates and those of Murata et al. (2004). This set coincides with those satisfying three common assumptions about surrogates. Equivalent expressions for the members-sometimes well known-follow for convex and concave surrogates, frequently used in the induction of linear separators and decision trees. Most notably, they share remarkable algorithmic features: for each of these two types of classifiers, we give a minimization algorithm provably converging to the minimum of any such surrogate. While seemingly different, we show that these algorithms are offshoots of the same ldquomasterrdquo algorithm. This provides a new and broad unified account of different popular algorithms, including additive regression with the squared loss, the logistic loss, and the top-down induction performed in CART, C4.5. Moreover, we show that the induction enjoys the most popular boosting features, regardless of the surrogate. Experiments are provided on 40 readily available domains.

Semi-supervised statistical region refinement for color image segmentation

Some authors have recently devised adaptations of spectral grouping algorithms to integrate prior knowledge, as constrained eigenvalues problems. In this paper, we improve and adapt a recent statistical region merging approach to this task, as a non-parametric mixture model estimation problem. The approach appears to be attractive both for its theoretical benefits and its experimental results, as slight bias brings dramatic improvements over unbiased approaches on challenging digital pictures.

Mixed Bregman Clustering with Approximation Guarantees

Two recent breakthroughs have dramatically improved the scope and performance of k-means clustering: squared Euclidean seeding for the initialization step, and Bregman clustering for the iterative step. In this paper, we first unite the two frameworks by generalizing the former improvement to Bregman seeding-- a biased randomized seeding technique using Bregman divergences -- while generalizing its important theoretical approximation guarantees as well. We end up with a complete Bregman hard clustering algorithm integrating the distortion at hand in both the initialization and iterative steps. Our second contribution is to further generalize this algorithm to handle mixed Bregman distortions, which smooth out the asymetricity of Bregman divergences. In contrast to some other symmetrization approaches, our approach keeps the algorithm simple and allows us to generalize theoretical guarantees from regular Bregman clustering. Preliminary experiments show that using the proposed seeding with a suitable Bregman divergence can help us discover the underlying structure of the data.

Fitting the smallest enclosing bregman ball

Finding a point which minimizes the maximal distortion with respect to a dataset is an important estimation problem that has recently received growing attentions in machine learning, with the advent of one class classification. We propose two theoretically founded generalizations to arbitrary Bregman divergences, of a recent popular smallest enclosing ball approximation algorithm for Euclidean spaces coined by Bădoiu and Clarkson in 2002.