Contribution to supervised representation learning: algorithms and applications.

Abstract

In this thesis, we focus on supervised learning methods for pattern categorization. In this context, itremains a major challenge to establish efficient relationships between the discriminant properties of theextracted features and the inter-class sparsity structure.Our first attempt to address this problem was to develop a method called "Robust Discriminant Analysiswith Feature Selection and Inter-class Sparsity" (RDA_FSIS). This method performs feature selectionand extraction simultaneously. The targeted projection transformation focuses on the most discriminativeoriginal features while guaranteeing that the extracted (or transformed) features belonging to the sameclass share a common sparse structure, which contributes to small intra-class distances.In a further study on this approach, some improvements have been introduced in terms of theoptimization criterion and the applied optimization process. In fact, we proposed an improved version ofthe original RDA_FSIS called "Enhanced Discriminant Analysis with Class Sparsity using GradientMethod" (EDA_CS). The basic improvement is twofold: on the first hand, in the alternatingoptimization, we update the linear transformation and tune it with the gradient descent method, resultingin a more efficient and less complex solution than the closed form adopted in RDA_FSIS.On the other hand, the method could be used as a fine-tuning technique for many feature extractionmethods. The main feature of this approach lies in the fact that it is a gradient descent based refinementapplied to a closed form solution. This makes it suitable for combining several extraction methods andcan thus improve the performance of the classification process.In accordance with the above methods, we proposed a hybrid linear feature extraction scheme called"feature extraction using gradient descent with hybrid initialization" (FE_GD_HI). This method, basedon a unified criterion, was able to take advantage of several powerful linear discriminant methods. Thelinear transformation is computed using a descent gradient method. The strength of this approach is thatit is generic in the sense that it allows fine tuning of the hybrid solution provided by different methods.Finally, we proposed a new efficient ensemble learning approach that aims to estimate an improved datarepresentation. The proposed method is called "ICS Based Ensemble Learning for Image Classification"(EM_ICS). Instead of using multiple classifiers on the transformed features, we aim to estimate multipleextracted feature subsets. These were obtained by multiple learned linear embeddings. Multiple featuresubsets were used to estimate the transformations, which were ranked using multiple feature selectiontechniques. The derived extracted feature subsets were concatenated into a single data representationvector with strong discriminative properties.Experiments conducted on various benchmark datasets ranging from face images, handwritten digitimages, object images to text datasets showed promising results that outperformed the existing state-ofthe-art and competing methods.