O processamento de sinais feito pelos sistemas neurais biológicos é altamente eficiente e complexo, por isso desperta grande atenção de pesquisa. Basicamente, todo o processamento de sinais funciona com base em redes de neurônios que emitem e recebem pulsos. Portanto, de forma geral, os estímulos recebidos do sistema sensorial por uma rede neural biológica de algum modo são convertidos em trens de pulsos. Aqui, nesta tese, é apresentada uma nova arquitetura composta por duas camadas: a primeira recebe correntes de estímulos de entrada e os mapeia em trens de pulsos; a segunda recebe esses trens de pulsos e os clássica em conjuntos de estímulos. Na primeira camada, a conversão de correntes de estímulos em trens de pulso é feita através de uma rede de neurônios osciladores acoplados por pulso. Esses neurônios possuem uma frequência natural de disparo e quando são agrupados em redes podem se coordenar para apresentar uma dinâmica global a longo prazo. Por sua vez, a dinâmica global também é sensível às correntes de entrada. Na segunda camada, a classificação dos trens de pulsos em conjuntos de estímulos é implementada por um neurônio do tipo integra-e-dispara. O comportamento típico para esse neurônio é de disparar ao menos uma vez para todas as integrações de trens de pulsos de uma determinada classe; caso contrário, ele deve car em silêncio. O processo de aprendizado da segunda camada depende do conhecimento do intervalo de tempo de repetição de um trem de pulsos. Portanto, nesta tese, são apresentadas métricas para definir tal intervalo de tempo, dando, assim, autonomia para a arquitetura. É possível concluir com base nos ensaios realizados que a arquitetura desenvolvida possui uma grande capacidade para mapeamento de correntes de entradas em trens de pulsos sem a necessidade de alterações na estrutura da arquitetura; também que a adição da dimensão tempo pela primeira camada ajuda na classificação realizada pela segunda. Assim, um novo modelo para realizar processos de codificação e decodificação é apresentado, desenvolvido através de séries de ensaios computacionais e caracterizado por medidas de sua dinâmica. / The signal processing done by the neural systems is highly efficient and complex, so that it attracts a large attention for research. Basically, all the signal processing functions are based on networks of neurons that send and receive spikes. Therefore, in general, the stimuli received from the sensory system by a biological neural network somehow are converted into spike trains. Here, in this thesis, we present a new architecture composed of two layers: the first layer receives streams of input stimuli and maps them on spike trains; the second layer receives these spike trains and classifies them in a sets of stimuli. In the first layer, the conversion of currents of stimuli on spike trains is made by a pulse-coupled neural network. Neurons in this context are like oscillators and have a natural frequency to shoot; when they are grouped into networks, they can be coordinated to present a global long-term dynamics. In turn, this global dynamics is also sensible to the input currents. In the second layer, the classification of spike trains in sets of stimuli is implemented by an integrate-and-re neuron. The typical behavior for this neuron is to shoot at least once every time that it receives a known spike train; otherwise, it should be in silence. The learning process of the second layer depends on the knowledge of the time interval of repetition of a spike train. Therefore, in this thesis, metrics are presented to define this time interval, thus giving autonomy to the architecture. It can be concluded on the basis of the tests developed that the architecture has a large capacity for mapping input currents on spike trains without requiring changes in its structure; moreover, the addition of the time dimension done by the first layer helps in the classification performed by the second layer. Thus, a new model to perform the encoding and decoding processes is presented, developed through a series of computational experiments and characterized by measurements of its dynamics.

O processamento de sinais feito pelos sistemas neurais biológicos é altamente eficiente e complexo, por isso desperta grande atenção de pesquisa. Basicamente, todo o processamento de sinais funciona com base em redes de neurônios que emitem e recebem pulsos. Portanto, de forma geral, os estímulos recebidos do sistema sensorial por uma rede neural biológica de algum modo são convertidos em trens de pulsos. Aqui, nesta tese, é apresentada uma nova arquitetura composta por duas camadas: a primeira recebe correntes de estímulos de entrada e os mapeia em trens de pulsos; a segunda recebe esses trens de pulsos e os clássica em conjuntos de estímulos. Na primeira camada, a conversão de correntes de estímulos em trens de pulso é feita através de uma rede de neurônios osciladores acoplados por pulso. Esses neurônios possuem uma frequência natural de disparo e quando são agrupados em redes podem se coordenar para apresentar uma dinâmica global a longo prazo. Por sua vez, a dinâmica global também é sensível às correntes de entrada. Na segunda camada, a classificação dos trens de pulsos em conjuntos de estímulos é implementada por um neurônio do tipo integra-e-dispara. O comportamento típico para esse neurônio é de disparar ao menos uma vez para todas as integrações de trens de pulsos de uma determinada classe; caso contrário, ele deve car em silêncio. O processo de aprendizado da segunda camada depende do conhecimento do intervalo de tempo de repetição de um trem de pulsos. Portanto, nesta tese, são apresentadas métricas para definir tal intervalo de tempo, dando, assim, autonomia para a arquitetura. É possível concluir com base nos ensaios realizados que a arquitetura desenvolvida possui uma grande capacidade para mapeamento de correntes de entradas em trens de pulsos sem a necessidade de alterações na estrutura da arquitetura; também que a adição da dimensão tempo pela primeira camada ajuda na classificação realizada pela segunda. Assim, um novo modelo para realizar processos de codificação e decodificação é apresentado, desenvolvido através de séries de ensaios computacionais e caracterizado por medidas de sua dinâmica. / The signal processing done by the neural systems is highly efficient and complex, so that it attracts a large attention for research. Basically, all the signal processing functions are based on networks of neurons that send and receive spikes. Therefore, in general, the stimuli received from the sensory system by a biological neural network somehow are converted into spike trains. Here, in this thesis, we present a new architecture composed of two layers: the first layer receives streams of input stimuli and maps them on spike trains; the second layer receives these spike trains and classifies them in a sets of stimuli. In the first layer, the conversion of currents of stimuli on spike trains is made by a pulse-coupled neural network. Neurons in this context are like oscillators and have a natural frequency to shoot; when they are grouped into networks, they can be coordinated to present a global long-term dynamics. In turn, this global dynamics is also sensible to the input currents. In the second layer, the classification of spike trains in sets of stimuli is implemented by an integrate-and-re neuron. The typical behavior for this neuron is to shoot at least once every time that it receives a known spike train; otherwise, it should be in silence. The learning process of the second layer depends on the knowledge of the time interval of repetition of a spike train. Therefore, in this thesis, metrics are presented to define this time interval, thus giving autonomy to the architecture. It can be concluded on the basis of the tests developed that the architecture has a large capacity for mapping input currents on spike trains without requiring changes in its structure; moreover, the addition of the time dimension done by the first layer helps in the classification performed by the second layer. Thus, a new model to perform the encoding and decoding processes is presented, developed through a series of computational experiments and characterized by measurements of its dynamics.

No description available.

No description available.

Les systèmes « excitables » sont omniprésents dans la nature, le plus paradigmatique d'entre eux étant le neurone, qui répond de façon « tout ou rien » aux perturbations externes. Cette particularité étant clairement établie comme l'un des points clé pour le fonctionnement des systèmes nerveux, son analyse dans des systèmes modèles (mathématiques ou physiques) peut d'une part aider à la compréhension de la dynamique d'ensembles de neurones couplés et d'autre part ouvrir des voies pour un traitement neuromimétique de l'information. C'est dans cette logique que s'inscrit la préparation de cette thèse de doctorat. Dans ce mémoire, nous utilisons des systèmes basés sur des lasers à semiconducteur pour d'une part modéliser des systèmes excitables ou des ensembles de systèmes neuromimétiques couplés et d'autre part pour contrôler (grâce à l'optogénétique) des canaux ioniques impliqués dans l'émission de potentiels d'action par des neurones de mammifères. Le long du premier chapitre, nous présentons de manière synthétique les concepts dynamiques sur lesquels nous nous appuierons dans la suite du manuscrit. Par la suite, nous décrivons brièvement le contexte de ce travail du point de vue de la synchronisation, notamment de cellules excitables. Enfin, nous discutons le contexte applicatif potentiel de ces travaux, c’est-à-dire l'utilisation de systèmes photoniques dits « neuromimétiques » dans le but de traiter de l'information. Dans le chapitre 2, nous analysons tout d'abord du point de vue théorique et bibliographique le caractère excitable d'un laser à semiconducteur sous l'influence d'un forçage optique cohérent. Par la suite, nous détaillons nos travaux expérimentaux d'abord, puis numériques et théoriques, sur la réponse de ce système « neuromimétique » à des perturbations répétées dans le temps. Tandis que le modèle mathématique simplifié prévoit un comportement de type intégrateur en réponse a des perturbations répétées, nous montrons que le comportement est en fait souvent résonateur, ce qui confère à ce système la propriété étonnante d'émettre une impulsion seulement s'il reçoit deux perturbations séparées d'un intervalle de temps bien précis. Nous montrons également que ce système peut convertir des perturbations de différente intensité en une série d'impulsions toutes identiques mais dont le nombre dépend de l'intensité de la perturbation incidente. Dans le chapitre 3, nous analysons (de nouveau expérimentalement, puis numériquement et théoriquement) le comportement dynamique d'un réseau de lasers à semiconducteur couplés dans un régime de chaos lent-rapide. Nous nous basons sur une étude antérieure montrant qu'un seul de ces éléments peut présenter une dynamique neuromimétique (en particulier l'émission chaotique d'impulsions originant du phénomène de canard). De façon surprenante pour un système ayant un si grand nombre de degrés de liberté, nous observons une dynamique qui semble chaotique de basse dimension. Nous examinons l'impact des propriétés statistiques de la population considérée sur la dynamique et relions nos observations expérimentales et numériques à l'existence d'une variété critique calculable analytiquement pour le champ moyen et près duquel converge la dynamique grâce au caractère lent-rapide du système. Dans le chapitre 4 enfin, nous présentons une brève étude expérimentale de la réponse de cellules biologiques à des perturbations lumineuses. En effet, les techniques optogénétiques permettent de rendre des cellules (en particulier des neurones) sensibles à la lumière grâce au contrôle optique de l'ouverture et de la fermeture de canaux ioniques. Ainsi, après avoir étudié dans les chapitres précédents des systèmes optiques sur la base de considérations provenant de systèmes biologiques, nous amenons matériellement un système laser vers un système biologique. / Excitable systems are everywhere in Nature, and among them the neuron, which responds to an external stimulus with an all-or-none type of response, is often regarded as the most typical example. This excitability behaviour is clearly established as to be one of the underlying operating mechanisms of the nervous system and its analysis in model systems (being them mathematical of physical) can, from one hand, shed some light on the dynamics of neural networks, and from the other, open novel ways for a neuro-mimetic treatment of information. The work presented in this PhD thesis was realized in this perspective. In this dissertation we will consider systems based on semiconductor lasers both for modelling excitable systems or coupled neuromorphic networks and for controlling (in an optogenetic outlook) ionic channels that are involved in the emission of action potentials of neurons in mammals. During the first chapter, we will briefly present the dynamical concepts on which we will build our understanding for the rest of the manuscript. Thereafter, we will describe the context of this work from the point of view of synchronized systems, in particular excitable cells. Finally, we will discuss in this context the applications potential of this work, namely the possibility of using “neuromimetic” photonic systems as a was to treat information. In chapter 2 we will firstly analyse from a theoretical and bibliographical standpoint the excitable character of a laser with coherent injection. Later, we will firstly detail our results, firstly experimental and subsequently numerical and theoretical, on the response of this “neuromimetic” system to perturbations repeated in time. Whereas the simplified mathematical model envisions an integrator behaviour in response to repeated perturbations, we will show that the system often acts as a resonator, thus imparting the remarkable property of being able to emit a single pulse only if it receives two perturbations that are separated by a specific time interval. We will also illustrate how this system can convert perturbations of different intensity in a series of all identical pulses whose number depends on the intensity of the incoming perturbation. In the third chapter we will analyse, first experimentally and later numerically and theoretically, the dynamical behaviour of a network of coupled semiconductor lasers in a slow-fast chaotic regime. We will rely on a previous study documenting that a single such element can present a neuromimetic dynamics (in particular, the emission of chaotic pulses originating from a canard phenomenon). Surprisingly for a system having such a large number of degrees of freedom, we observe a dynamics which seems low dimensional chaotic. We will examine the impact of statistical properties of the selected population on the dynamics, and we will link our experimental and numerical observations to the existence of a slow manifold for the mean field, computable analytically, and towards whom the dynamics converges thanks to the slow-fact nature of the system. Finally, in chapter 4 we will present a short experimental study on the response of biological cells to light perturbations. Indeed, optogenetic techniques enables to render the cells (in particular neurons) sensitive to light due to the optical control of the opening and closing of ionic channels. Hence, after having studied in the previous chapters optical systems on the basis of observations derived from biological systems, we will physically transfer an optical system towards a biological one. Here we lay the groundwork of a photonic system which allows, with a moderate complexity, to realize cell measurements in response to spatially localized optical perturbations.