The thesis deals with the analysis of the clustering and mapping techniques derived from the principles of the neural and statistical learning and growth theory. The selected branch of the unsupervised bio-inspired prototype-based models is described in terms of the proposed logical framework, which highlights the continuity of these methods with the classical "pure" statistical methods. Moreover, as those methods are broadly understood as the "black boxes" with the unpredictable, unclear and especially hidden behavior, the examples of the spatial computational and organizational patterns in two-dimensional space are provided. Additionally, this thesis presents the novel concept based on the non-linear, non-Gaussian Gompertzian function, which has been widely used as the universal law in dynamic growth models, but has not yet been applied in the field of computational intelligence. The essence of Gompertzian dynamics is mathematically analyzed and a novel simple version of the Gompertzian normalized function is introduced. Furthermore, the function was modified for use in the field of artificial intelligence and neural implications were discussed. Additionally, the novel neural networks were proposed and derived from the topological principles of Kohonen's self-organizing maps and neural gas algorithm. The Gompertzian networks were evaluated using several indicators for various generated and real datasets. Gompertzian neural networks with fixed grid and integrated neighborhood ranking principle generally show lower mean squared errors than the original SOM algorithms. Likewise, the unconstrained Gompertzian networks have demonstrated overall low error rates comparable to neural gas algorithm, more stable and lower error solutions than the k- means sequential procedure. In conclusion, the Gompertzian function has been shown to be a viable concept and an effective computational tool for multidimensional data analysis.

Cette thèse s'intéresse à la problématique de la classification automatique des séquences vidéo. L'idée est de se démarquer de la méthodologie dominante qui se base sur l'utilisation de caractéristiques conçues manuellement, et de proposer des modèles qui soient les plus génériques possibles et indépendants du domaine. Ceci est fait en automatisant la phase d'extraction des caractéristiques, qui sont dans notre cas générées par apprentissage à partir d'exemples, sans aucune connaissance a priori. Nous nous appuyons pour ce faire sur des travaux existants sur les modèles neuronaux pour la reconnaissance d'objets dans les images fixes, et nous étudions leur extension au cas de la vidéo. Plus concrètement, nous proposons deux modèles d'apprentissage des caractéristiques spatio-temporelles pour la classification vidéo : (i) Un modèle d'apprentissage supervisé profond, qui peut être vu comme une extension des modèles ConvNets au cas de la vidéo, et (ii) Un modèle d'apprentissage non supervisé, qui se base sur un schéma d'auto-encodage, et sur une représentation parcimonieuse sur-complète des données. Outre les originalités liées à chacune de ces deux approches, une contribution supplémentaire de cette thèse est une étude comparative entre plusieurs modèles de classification de séquences parmi les plus populaires de l'état de l'art. Cette étude a été réalisée en se basant sur des caractéristiques manuelles adaptées à la problématique de la reconnaissance d'actions dans les vidéos de football. Ceci a permis d'identifier le modèle de classification le plus performant (un réseau de neurone récurrent bidirectionnel à longue mémoire à court-terme -BLSTM-), et de justifier son utilisation pour le reste des expérimentations. Enfin, afin de valider la généricité des deux modèles proposés, ceux-ci ont été évalués sur deux problématiques différentes, à savoir la reconnaissance d'actions humaines (sur la base KTH), et la reconnaissance d'expressions faciales (sur la base GEMEP-FERA). L'étude des résultats a permis de valider les approches, et de montrer qu'elles obtiennent des performances parmi les meilleures de l'état de l'art (avec 95,83% de bonne reconnaissance pour la base KTH, et 87,57% pour la base GEMEP-FERA). / This thesis focuses on the issue of automatic classification of video sequences. We aim, through this work, at standing out from the dominant methodology, which relies on so-called hand-crafted features, by proposing generic and problem-independent models. This can be done by automating the feature extraction process, which is performed in our case through a learning scheme from training examples, without any prior knowledge. To do so, we rely on existing neural-based methods, which are dedicated to object recognition in still images, and investigate their extension to the video case. More concretely, we introduce two learning-based models to extract spatio-temporal features for video classification: (i) A deep learning model, which is trained in a supervised way, and which can be considered as an extension of the popular ConvNets model to the video case, and (ii) An unsupervised learning model that relies on an auto-encoder scheme, and a sparse over-complete representation. Moreover, an additional contribution of this work lies in a comparative study between several sequence classification models. This study was performed using hand-crafted features especially designed to be optimal for the soccer action recognition problem. Obtained results have permitted to select the best classifier (a bidirectional long short-term memory recurrent neural network -BLSTM-) to be used for all experiments. In order to validate the genericity of the two proposed models, experiments were carried out on two different problems, namely human action recognition (using the KTH dataset) and facial expression recognition (using the GEMEP-FERA dataset). Obtained results show that our approaches achieve outstanding performances, among the best of the related works (with a recognition rate of 95,83% for the KTH dataset, and 87,57% for the GEMEP-FERA dataset).

One of the most powerful and important applications that the brain accomplishes is solving the sensory "cocktail party problem:" to adaptively suppress extraneous signals in an environment. Theoretical studies suggest that the solution to the problem involves an adaptive filter, which learns to remove the redundant noise. However, neural learning is also in its infancy and there are still many questions about the stability and application of synaptic learning rules for neural computation. In this thesis, the implementation of an adaptive filter in the brain of a weakly electric fish, A. Leptorhynchus, was studied. It was found to require a cerebellar architecture that could supply independent frequency channels of delayed feedback and multiple burst learning rules that could shape this feedback. This unifies two ideas about the function of the cerebellum that were previously separate: the cerebellum as an adaptive filter and as a generator of precise temporal inputs.

One of the most powerful and important applications that the brain accomplishes is solving the sensory "cocktail party problem:" to adaptively suppress extraneous signals in an environment. Theoretical studies suggest that the solution to the problem involves an adaptive filter, which learns to remove the redundant noise. However, neural learning is also in its infancy and there are still many questions about the stability and application of synaptic learning rules for neural computation. In this thesis, the implementation of an adaptive filter in the brain of a weakly electric fish, A. Leptorhynchus, was studied. It was found to require a cerebellar architecture that could supply independent frequency channels of delayed feedback and multiple burst learning rules that could shape this feedback. This unifies two ideas about the function of the cerebellum that were previously separate: the cerebellum as an adaptive filter and as a generator of precise temporal inputs.

One of the most powerful and important applications that the brain accomplishes is solving the sensory "cocktail party problem:" to adaptively suppress extraneous signals in an environment. Theoretical studies suggest that the solution to the problem involves an adaptive filter, which learns to remove the redundant noise. However, neural learning is also in its infancy and there are still many questions about the stability and application of synaptic learning rules for neural computation. In this thesis, the implementation of an adaptive filter in the brain of a weakly electric fish, A. Leptorhynchus, was studied. It was found to require a cerebellar architecture that could supply independent frequency channels of delayed feedback and multiple burst learning rules that could shape this feedback. This unifies two ideas about the function of the cerebellum that were previously separate: the cerebellum as an adaptive filter and as a generator of precise temporal inputs.

One of the most powerful and important applications that the brain accomplishes is solving the sensory "cocktail party problem:" to adaptively suppress extraneous signals in an environment. Theoretical studies suggest that the solution to the problem involves an adaptive filter, which learns to remove the redundant noise. However, neural learning is also in its infancy and there are still many questions about the stability and application of synaptic learning rules for neural computation. In this thesis, the implementation of an adaptive filter in the brain of a weakly electric fish, A. Leptorhynchus, was studied. It was found to require a cerebellar architecture that could supply independent frequency channels of delayed feedback and multiple burst learning rules that could shape this feedback. This unifies two ideas about the function of the cerebellum that were previously separate: the cerebellum as an adaptive filter and as a generator of precise temporal inputs.