This thesis addresses the problem of how the dendritic structure and other morphological properties of the neuron can determine its pattern recognition performance. The techniques used in this work for generating dendritic trees with different morphologies included the following three methods. Firstly, dendritic trees were produced by exhaustively generating every possible morphology. Where this was not possible due to the size of morphological space, I sampled systematically from the possible morphologies. Lastly, dendritic trees were evolved using an evolutionary algorithm, which varied existing morphologies using selection, mutation and crossover. From these trees, I constructed full compartmental conductance-based models of neurons. I then assessed the performance of the resulting neuronal models by quantifying their ability to discriminate between learned and novel input patterns. The morphologies generated were tested in the presence and absence of active conductances. The results have shown that the morphology does have a considerable effect on pattern recognition performance. In fact, neurons with a small mean depth of their dendritic tree are the best pattern recognizers. Moreover, the performance of neurons is anti-correlated with mean depth. Interestingly, the symmetry of the neuronal morphology does not correlate with performance. This research has also revealed that the evolutionary algorithm could find effective morphologies for both passive models and models with active conductances. In the active model, there was a considerable change in the performance of the original population of neurons, which largely resulted from changes in the morphological parameters such as dendritic compartmental length and tapering. However, no single parameter setting guaranteed good neuronal performance; in three separate runs of the evolutionary algorithm, different sets of well performing parameters were found. In fact, the evolved neurons performed at least five times better than the original hand-tuned neurons. In summary, the combination of morphological parameters plays a key role in determining the performance of neurons in the pattern recognition task and the right combination produces very well performing neurons.

According to the classical efficient-coding hypothesis, biological neurons are naturally adapted to transmit and process information about the stimulus in an optimal way. Shannon's information theory provides methods to compute the fundamental limits on maximal information transfer by a general system. Understanding how these limits differ between different classes of neurons may help us to better understand how sensory and other information is processed in the brain. In this work we provide a brief review of information theory and its use in computational neuroscience. We use mathematical models of neuronal cells with stochastic input that realistically reproduce different activity patterns observed in real cortical neurons. By employing the neuronal input-output pro- perties we calculate several key information-theoretic characteristics, including the information capacity. In order to determine the information capacity we propose an iterative extension of the Blahut-Arimoto algorithm that generalizes to continuous input channels subjected to constraints. Finally, we compare the information optimality conditions among different models and parameter sets. 1

Nous proposons un algorithme basé sur une approche orientée d'ensemble de système dynamique pour extraire une organisation grossière de l'espace d'état de cerveau sur la base des signaux de l'EEG. Nous l'utilisons pour comparer l'organisation de l'espace d'état des données simulées à grande échelle avec la dynamique cérébrale réelle au repos chez des sujets sains et pathologiques (SEP). / We propose an algorithme based on set oriented approach of dynamical system to extract a coarse grained organization of brain state space on the basis of EEG signals. We use it for comparing the organization of the state space of large scale simulation of brain dynamics with actual brain dynamics of resting activity in healthy and SEP subjects.

In this work is presented stochastic model of binaural hearing in context of another alternative models. According to latest experimental data on mammals, inhibition plays a role in interaural time difference recognition, which is a key for low frequency sound source localization. The outputs of experiments may lead to the conclusion that the binaural hearing works differently in mammals compared to birds. Nowadays there are a few theoretical works addressing this new phenomena, but all of them are relaying on a very precise inhibition timing, which was never proved as physiologically valid. On the other hand, models described in this work are based on the fact, that every neuron has a random delay when reacting to an excitation. If this time jitter is taken into account and combined with inhibitory signal, delay in the neuronal circuit and coincidence detection, then the output firing rate corresponds to the azimuth of the sound source. In this work it is shown, that such a neuronal circuits are giving the same output results compared to experimental data. The models are supported by analytical computations and numerical simulations including simulation of cochlear implant.

In this work is presented stochastic model of binaural hearing in context of another alternative models. According to latest experimental data on mammals, inhibition plays a role in interaural time difference recognition, which is a key for low frequency sound source localization. The outputs of experiments may lead to the conclusion that the binaural hearing works differently in mammals compared to birds. Nowadays there are a few theoretical works addressing this new phenomena, but all of them are relaying on a very precise inhibition timing, which was never proved as physiologically valid. On the other hand, models described in this work are based on the fact, that every neuron has a random delay when reacting to an excitation. If this time jitter is taken into account and combined with inhibitory signal, delay in the neuronal circuit and coincidence detection, then the output firing rate corresponds to the azimuth of the sound source. In this work it is shown, that such a neuronal circuits are giving the same output results compared to experimental data. The models are supported by analytical computations and numerical simulations including simulation of cochlear implant.

L’utilisation de techniques numériques lors de la conception d’un produit s’est largement généralisée au cours des 30 dernières années. Pourtant, la lenteur des calculs et la spécialisation des modèles numériques restent problématiques.Nous avons donc choisi de développer des modèles réduits du comportement de balayage sur un moteur Diesel 2-temps à lumières. Ces modèles sont analytiques, génériques, rapides d’utilisation et permettent d’éliminer les problématiques de traitement numérique. Ils sont aussi des instruments performants dans la recherche de solutions de conception. Une modélisation CFD 2D a tout d’abord été développée pour servir de bases de données, avec la définition des paramètres primordiaux à suivre pour quantifier un balayage optimal.Le travail de recherche a dévoilé une méthodologie nouvelle fondée sur un méta-modèle du comportement dit « neuro-séparé » comprenant un modèle neuronal d’état, un modèle neuronal pseudo-dynamique et un modèle à variables séparées. Ensuite, un processus d'aide à la décision exploitant les modèles précédents a été mis en place au travers d’un processus d’optimisation évolutionnaire (fondé sur les algorithmes génétiques) puis de la simulation comportementale rapide des solutions optimales de conception par un krigeage.La démarche de conception multipoints de vue, multi-critères et multi-physiques appliquée au moteur intègre aussi une dimension cognitive : l’exploration évolutionnaire des espaces de solutions a été menée de façon libre et forcée. Afin de valider notre approche, nous avons mis en place des critères de qualification appliqués à chacun de nos modèles, permettant de quantifier les écarts visà-vis de la base initiale CFD qui a fondé nos modèles réduits.Notre démarche a mené à la création d’un outil d’aide à la modélisation et à la décision exploitant les modules Python et Matlab développés. / The use of numerical methods to design a product became more and more commonover the past 30 years. However, numerical models are still specialized and they do not run fastwhich make their use problematic. So some reduced models of scavenging have been developed. These models are analytical andgeneric; they run quickly and avoid the numerical treatment problems. They are also some efficienttools in the search of design solutions.The work carried out has led to a new methodology based on a behavioral meta-model called“neuro-separated” including a neuronal model of state, a pseudo-dynamic neuronal model and amodel with separated variables. Then, a process of decision aids exploiting the models previouslydeveloped in evolutionary algorithms (genetic algorithms) and the fast behavioral simulation of theoptimal design solutions thanks to the kriging approach.This design approach is multi-viewpoints, multi-criteria and multi-physics. It also includes acognitive dimension: both free and controlled evolutionary explorations of solution spaces have beendone. To validate the method, some qualification criteria have been evaluated for each model. Theyallow to understand and to assume the gap between the reduced models and the initial CFD base(where the model are coming from). Our approach has led to the development of a tool of model and decision aids using Python and Matlab software programs.