Au cours de la dernière décennie, l'apprentissage profond est devenu une composante majeure de l'intelligence artificielle, ayant mené à une série d'avancées capitales dans une variété de domaines. L'un des piliers de l'apprentissage profond est l'optimisation de fonction de coût par l'algorithme du gradient stochastique (SGD). Traditionnellement en apprentissage profond, les réseaux de neurones sont des fonctions mathématiques différentiables, et les gradients requis pour l'algorithme SGD sont calculés par rétropropagation. Cependant, les architectures informatiques sur lesquelles ces réseaux de neurones sont implémentés et entraînés souffrent d’inefficacités en vitesse et en énergie, dues à la séparation de la mémoire et des calculs dans ces architectures. Pour résoudre ces problèmes, le neuromorphique vise à implementer les réseaux de neurones dans des architectures qui fusionnent mémoire et calculs, imitant plus fidèlement le cerveau. Dans cette thèse, nous soutenons que pour construire efficacement des réseaux de neurones dans des architectures neuromorphiques, il est nécessaire de repenser les algorithmes pour les implémenter et les entraîner. Nous présentons un cadre mathématique alternative, compatible lui aussi avec l’algorithme SGD, qui permet de concevoir des réseaux de neurones dans des substrats qui exploitent mieux les lois de la physique. Notre cadre mathématique s'applique à une très large classe de modèles, à savoir les systèmes dont l'état ou la dynamique sont décrits par des équations variationnelles. La procédure pour calculer les gradients de la fonction de coût dans de tels systèmes (qui dans de nombreux cas pratiques ne nécessite que de l'information locale pour chaque paramètre) est appelée “equilibrium propagation” (EqProp). Comme beaucoup de systèmes en physique et en ingénierie peuvent être décrits par des principes variationnels, notre cadre mathématique peut potentiellement s'appliquer à une grande variété de systèmes physiques, dont les applications vont au delà du neuromorphique et touchent divers champs d'ingénierie. / In the last decade, deep learning has become a major component of artificial intelligence, leading to a series of breakthroughs across a wide variety of domains. The workhorse of deep learning is the optimization of loss functions by stochastic gradient descent (SGD). Traditionally in deep learning, neural networks are differentiable mathematical functions, and the loss gradients required for SGD are computed with the backpropagation algorithm. However, the computer architectures on which these neural networks are implemented and trained suffer from speed and energy inefficiency issues, due to the separation of memory and processing in these architectures. To solve these problems, the field of neuromorphic computing aims at implementing neural networks on hardware architectures that merge memory and processing, just like brains do. In this thesis, we argue that building large, fast and efficient neural networks on neuromorphic architectures also requires rethinking the algorithms to implement and train them. We present an alternative mathematical framework, also compatible with SGD, which offers the possibility to design neural networks in substrates that directly exploit the laws of physics. Our framework applies to a very broad class of models, namely those whose state or dynamics are described by variational equations. This includes physical systems whose equilibrium state minimizes an energy function, and physical systems whose trajectory minimizes an action functional (principle of least action). We present a simple procedure to compute the loss gradients in such systems, called equilibrium propagation (EqProp), which requires solely locally available information for each trainable parameter. Since many models in physics and engineering can be described by variational principles, our framework has the potential to be applied to a broad variety of physical systems, whose applications extend to various fields of engineering, beyond neuromorphic computing.

This biomedical engineering thesis explores the opportunities and challenges of 2D+t contrast echocardiography for left ventricle functional analysis, both clinically and within a computer vision atlas-based deformable template model framework. A database was created for the experiments in this thesis, with 21 studies of contrast Dobutamine Stress Echo, in all 4 principal planes. The database includes clinical variables, human expert hand-traced myocardial contours and visual scoring. First the problem is studied from a clinical perspective. Quantification of endocardial global and local function using standard measures shows expected values and agreement with human expert visual scoring, but the results are less reliable for myocardial thickening. Next, the problem of segmenting the endocardium with a computer is posed in a standard landmark and atlas-based deformable template model framework. The underlying assumption is that these models can emulate human experts in terms of integrating previous knowledge about the anatomy and physiology with three sources of information from the image: texture, geometry and kinetics. Probabilistic atlases of contrast echocardiography are computed, while noting from histograms at selected anatomical locations that modelling texture with just mean intensity values may be too naive. Intensity analysis together with the clinical results above suggest that lack of external boundary definition may preclude this imaging technique for appropriate measuring of myocardial thickening, while endocardial boundary definition is appropriate for evaluation of wall motion. Geometry is presented in a Principal Component Analysis (PCA) context, highlighting issues about Gaussianity, the correlation and covariance matrices with respect to physiology, and analysing different measures of dimensionality. A popular extension of deformable models ---Active Appearance Models (AAMs)--- is then studied in depth. Contrary to common wisdom, it is contended that using a PCA texture space instead of a fixed atlas is detrimental to segmentation, and that PCA models are not convenient for texture modelling. To integrate kinetics, a novel spatio-temporal model of cardiac contours is proposed. The new explicit model does not require frame interpolation, and it is compared to previous implicit models in terms of approximation error when the shape vector changes from frame to frame or remains constant throughout the cardiac cycle. Finally, the 2D+t atlas-based deformable model segmentation problem is formulated and solved with a gradient descent approach. Experiments using the similarity transformation suggest that segmentation of the whole cardiac volume outperforms segmentation of individual frames. A relatively new approach ---the inverse compositional algorithm--- is shown to decrease running times of the classic Lucas-Kanade algorithm by a factor of 20 to 25, to values that are within real-time processing reach.