Abstract: Deep learning has recently been enjoying an increasing popularity due to its success in solving challenging tasks. In particular, deep learning has proven to be effective in a large variety of computer vision tasks, such as image classification, object recognition and image parsing. Contrary to previous research, which required engineered feature representations, designed by experts, in order to succeed, deep learning attempts to learn representation hierarchies automatically from data. More recently, the trend has been to go deeper with representation hierarchies.
Learning (very) deep representation hierarchies is a challenging task, which
involves the optimization of highly non-convex functions. Therefore, the search
for algorithms to ease the learning of (very) deep representation hierarchies from data is extensive and ongoing.
In this thesis, we tackle the challenging problem of easing the learning of (very) deep representation hierarchies. We present a hyper-parameter free, off-the-shelf, simple and fast unsupervised algorithm to discover hidden structure from the input data by enforcing a very strong form of sparsity. We study the applicability and potential of the algorithm to learn representations of varying depth in a handful of applications and domains, highlighting the ability of the algorithm to provide discriminative feature representations that are able to achieve top performance.
Yet, while emphasizing the great value of unsupervised learning methods when
labeled data is scarce, the recent industrial success of deep learning has revolved around supervised learning. Supervised learning is currently the focus of many recent research advances, which have shown to excel at many computer vision tasks. Top performing systems often involve very large and deep models, which are not well suited for applications with time or memory limitations. More in line with the current trends, we engage in making top performing models more efficient, by designing very deep and thin models. Since training such very deep models still appears to be a challenging task, we introduce a novel algorithm that guides the training of very thin and deep models by hinting their intermediate representations.
Very deep and thin models trained by the proposed algorithm end up extracting feature representations that are comparable or even better performing
than the ones extracted by large state-of-the-art models, while compellingly
reducing the time and memory consumption of the model.

Abstract: This paper introduces the use of single-layer and deep convolutional networks for remote sensing data analysis. Direct application to multi- and hyperspectral imagery of supervised (shallow or deep) convolutional networks is very challenging given the high input data dimensionality and the relatively small amount of available labeled data. Therefore, we propose the use of greedy layerwise unsupervised pretraining coupled with a highly efficient algorithm for unsupervised learning of sparse features. The algorithm is rooted on sparse representations and enforces both population and lifetime sparsity of the extracted features, simultaneously. We successfully illustrate the expressive power of the extracted representations in several scenarios: classification of aerial scenes, as well as land-use classification in very high resolution or land-cover classification from multi- and hyperspectral images. The proposed algorithm clearly outperforms standard principal component analysis (PCA) and its kernel counterpart (kPCA), as well as current state-of-the-art algorithms of aerial classification, while being extremely computationally efficient at learning representations of data. Results show that single-layer convolutional networks can extract powerful discriminative features only when the receptive field accounts for neighboring pixels and are preferred when the classification requires high resolution and detailed results. However, deep architectures significantly outperform single-layer variants, capturing increasing levels of abstraction and complexity throughout the feature hierarchy.

Abstract: While depth tends to improve network performances, it also makes gradient-based training more difficult since deeper networks tend to be more non-linear. The recently proposed knowledge distillation approach is aimed at obtaining small and fast-to-execute models, and it has shown that a student network could imitate the soft output of a larger teacher network or ensemble of networks. In this paper, we extend this idea to allow the training of a student that is deeper and thinner than the teacher, using not only the outputs but also the intermediate representations learned by the teacher as hints to improve the training process and final performance of the student. Because the student intermediate hidden layer will generally be smaller than the teacher's intermediate hidden layer, additional parameters are introduced to map the student hidden layer to the prediction of the teacher hidden layer. This allows one to train deeper students that can generalize better or run faster, a trade-off that is controlled by the chosen student capacity. For example, on CIFAR-10, a deep student network with almost 10.4 times less parameters outperforms a larger, state-of-the-art teacher network.

Abstract: CoRR abs/1402.5766
We propose a meta-parameter free, off-the-shelf, simple and fast unsupervised feature learning algorithm, which exploits a new way of optimizing for sparsity. Experiments on STL-10 show that the method presents state-of-the-art performance and provides discriminative features that generalize well.

Abstract: Automatic face classification is currently a popular research area in Computer Vision. It involves several subproblems, such as subject recognition, gender classification or subject verification.
Current systems of automatic face classification need a large amount of training data to robustly learn a task. However, the collection of labeled data is usually a difficult issue. For this reason, the research on methods that are able to learn from a small sized training set is essential.
The dependency on the abundance of training data is not so evident in human learning processes. We are able to learn from a very small number of examples, given that we use, additionally, some prior knowledge to learn a new task. For example, we frequently find patterns and analogies from other domains to reuse them in new situations, or exploit training data from other experiences.
In computer science, Multitask Learning is a new Machine Learning approach that studies this idea of knowledge transfer among different tasks, to overcome the effects of the small sample sized problem.
This thesis explores, proposes and tests some Multitask Learning methods specially developed for face classification purposes. Moreover, it presents two more contributions dealing with the small sample sized problem, out of the Multitask Learning context. The first one is a method to extract external face features, to be used as an additional information source in automatic face classification problems. The second one is an empirical study on the most suitable face image resolution to perform automatic subject recognition.

Abstract: In this work we build a system for automatic emotion classification from image sequences. We analyze subtle changes in facial expressions by detecting a subset of 12 representative facial action units (AUs). Then, we classify emotions based on the output of these AUs classifiers, i.e. the presence/absence of AUs. We base the AUs classification upon a set of spatio-temporal geometric and appearance features for facial representation, fusing them within the emotion classifier. A decision tree is trained for emotion classifying, making the resulting model easy to interpret by capturing the combination of AUs activation that lead to a particular emotion. For Cohn-Kanade database, the proposed system classifies 7 emotions with a mean accuracy of near 90%, attaining a similar recognition accuracy in comparison with non-interpretable models that are not based in AUs detection.