Marc Sebban

Unsupervised Visual Domain Adaptation Using Subspace Alignment

In this paper, we introduce a new domain adaptation (DA) algorithm where the source and target domains are represented by subspaces described by eigenvectors. In this context, our method seeks a domain adaptation solution by learning a mapping function which aligns the source subspace with the target one. We show that the solution of the corresponding optimization problem can be obtained in a simple closed form, leading to an extremely fast algorithm. We use a theoretical result to tune the unique hyper parameter corresponding to the size of the subspaces. We run our method on various datasets and show that, despite its intrinsic simplicity, it outperforms state of the art DA methods.

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.

Learning stochastic edit distance: Application in handwritten character recognition

Many pattern recognition algorithms are based on the nearest-neighbour search and use the well-known edit distance, for which the primitive edit costs are usually fixed in advance. In this article, we aim at learning an unbiased stochastic edit distance in the form of a finite-state transducer from a corpus of (input, output) pairs of strings. Contrary to the other standard methods, which generally use the Expectation Maximisation algorithm, our algorithm learns a transducer independently on the marginal probability distribution of the input strings. Such an unbiased way to proceed requires to optimise the parameters of a conditional transducer instead of a joint one. We apply our new model in the context of handwritten digit recognition. We show, carrying out a large series of experiments, that it always outperforms the standard edit distance.

Discriminative feature fusion for image classification

Bag-of-words-based image classification approaches mostly rely on low level local shape features. However, it has been shown that combining multiple cues such as color, texture, or shape is a challenging and promising task which can improve the classification accuracy. Most of the state-of-the-art feature fusion methods usually aim to weight the cues without considering their statistical dependence in the application at hand. In this paper, we present a new logistic regression-based fusion method, called LRFF, which takes advantage of the different cues without being tied to any of them. We also design a new marginalized kernel by making use of the output of the regression model. We show that such kernels, surprisingly ignored so far by the computer vision community, are particularly well suited to achieve image classification tasks. We compare our approach with existing methods that combine color and shape on three datasets. The proposed learning-based feature fusion process clearly outperforms the state-of-the art fusion methods for image classification.

Learning probabilistic models of tree edit distance

Nowadays, there is a growing interest in machine learning and pattern recognition for tree-structured data. Trees actually provide a suitable structural representation to deal with complex tasks such as web information extraction, RNA secondary structure prediction, computer music, or conversion of semi-structured data (e.g. XML documents). Many applications in these domains require the calculation of similarities over pairs of trees. In this context, the tree edit distance (ED) has been subject of investigations for many years in order to improve its computational efficiency. However, used in its classical form, the tree ED needs a priori fixed edit costs which are often difficult to tune, that leaves little room for tackling complex problems. In this paper, to overcome this drawback, we focus on the automatic learning of a non-parametric stochastic tree ED. More precisely, we are interested in two kinds of probabilistic approaches. The first one builds a generative model of the tree ED from a joint distribution over the edit operations, while the second works from a conditional distribution providing then a discriminative model. To tackle these tasks, we present an adaptation of the expectation-maximization algorithm for learning these distributions over the primitive edit costs. Two experiments are conducted. The first is achieved on artificial data and confirms the interest to learn a tree ED rather than a priori imposing edit costs; The second is applied to a pattern recognition task aiming to classify handwritten digits.

Landmarks-based kernelized subspace alignment for unsupervised domain adaptation

Domain adaptation (DA) has gained a lot of success in the recent years in computer vision to deal with situations where the learning process has to transfer knowledge from a source to a target domain. In this paper, we introduce a novel unsupervised DA approach based on both subspace alignment and selection of landmarks similarly distributed between the two domains. Those landmarks are selected so as to reduce the discrepancy between the domains and then are used to non linearly project the data in the same space where an efficient subspace alignment (in closed-form) is performed. We carry out a large experimental comparison in visual domain adaptation showing that our new method outperforms the most recent unsupervised DA approaches.

Good edit similarity learning by loss minimization

Similarity functions are a fundamental component of many learning algorithms. When dealing with string or tree-structured data, measures based on the edit distance are widely used, and there exist a few methods for learning them from data. However, these methods offer no theoretical guarantee as to the generalization ability and discriminative power of the learned similarities. In this paper, we propose an approach to edit similarity learning based on loss minimization, called GESL. It is driven by the notion of (ϵ,γ,τ)-goodness, a theory that bridges the gap between the properties of a similarity function and its performance in classification. Using the notion of uniform stability, we derive generalization guarantees that hold for a large class of loss functions. We also provide experimental results on two real-world datasets which show that edit similarities learned with GESL induce more accurate and sparser classifiers than other (standard or learned) edit similarities.

Supervised learning of Gaussian mixture models for visual vocabulary generation

The creation of semantically relevant clusters is vital in bag-of-visual words models which are known to be very successful to achieve image classification tasks. Generally, unsupervised clustering algorithms, such as K-means, are employed to create such clusters from which visual dictionaries are deduced. K-means achieves a hard assignment by associating each image descriptor to the cluster with the nearest mean. By this way, the within-cluster sum of squares of distances is minimized. A limitation of this approach in the context of image classification is that it usually does not use any supervision that limits the discriminative power of the resulting visual words (typically the centroids of the clusters). More recently, some supervised dictionary creation methods based on both supervised information and data fitting were proposed leading to more discriminative visual words. But, none of them consider the uncertainty present at both image descriptor and cluster levels. In this paper, we propose a supervised learning algorithm based on a Gaussian mixture model which not only generalizes the K-means algorithm by allowing soft assignments, but also exploits supervised information to improve the discriminative power of the clusters. Technically, our algorithm aims at optimizing, using an EM-based approach, a convex combination of two criteria: the first one is unsupervised and based on the likelihood of the training data; the second is supervised and takes into account the purity of the clusters. We show on two well-known datasets that our method is able to create more relevant clusters by comparing its behavior with the state of the art dictionary creation methods.

A computerized prediction model of hazardous inflammatory platelet transfusion outcomes.

Background Platelet component (PC) transfusion leads occasionally to inflammatory hazards. Certain BRMs that are secreted by the platelets themselves during storage may have some responsibility. Methodology/Principal Findings First, we identified non-stochastic arrangements of platelet-secreted BRMs in platelet components that led to acute transfusion reactions (ATRs). These data provide formal clinical evidence that platelets generate secretion profiles under both sterile activation and pathological conditions. We next aimed to predict the risk of hazardous outcomes by establishing statistical models based on the associations of BRMs within the incriminated platelet components and using decision trees. We investigated a large (n = 65) series of ATRs after platelet component transfusions reported through a very homogenous system at one university hospital. Herein, we used a combination of clinical observations, ex vivo and in vitro investigations, and mathematical modeling systems. We calculated the statistical association of a large variety (n = 17) of cytokines, chemokines, and physiologically likely factors with acute inflammatory potential in patients presenting with severe hazards. We then generated an accident prediction model that proved to be dependent on the level (amount) of a given cytokine-like platelet product within the indicated component, e.g., soluble CD40-ligand (>289.5 pg/109 platelets), or the presence of another secreted factor (IL-13, >0). We further modeled the risk of the patient presenting either a febrile non-hemolytic transfusion reaction or an atypical allergic transfusion reaction, depending on the amount of the chemokine MIP-1α (<20.4 or >20.4 pg/109 platelets, respectively). Conclusions/Significance This allows the modeling of a policy of risk prevention for severe inflammatory outcomes in PC transfusion.

Platelet components associated with adverse reactions: predictive value of mitochondrial DNA relative to biological response modifiers

Biological response modifiers (BRMs), secreted by platelets (PLTs) during storage, play a role in adverse events (AEs) associated with transfusion. Moreover, mitochondrial DNA (mtDNA) levels in PLT components (PCs) are associated with AEs. In this study we explore whether there is a correlation between pathogenic BRMs and mtDNA levels and whether these markers can be considered predictors of transfusion pathology.