La robustesse des réseaux de neurones aux perturbations externes

par Allan Mancoo

Projet de thèse en Sciences cognitives option Neurosciences computationnelles

Sous la direction de Sophie Denève et de Christian Machens.

Thèses en préparation à Paris Sciences et Lettres , dans le cadre de Cerveau, cognition, comportement , en partenariat avec Laboratoire de Neurosciences Cognitives (laboratoire) et de École normale supérieure (Paris ; 1985-....) (établissement de préparation de la thèse) depuis le 01-09-2016 .


  • Résumé

    Il est observé que le cerveau est résilient contre la mort des neurones. En particulier, chez les patients qui souffrent de la maladie d'Alzheimer, plusieurs années se passent où la mort cellulaire s'établie avant qu'ils ne soient diagnostiqués cliniquement. Alors la question qu'on se pose est comment est-ce que les réseaux de neurones arrivent à compenser pour cette perte de neurones sans qu'il y ait de changements brusques dans le comportement de l'individu. Dans ce projet de recherche, on apporte une approche théorique où on fait l'hypothèse que dès qu'un certain groupe de neurones est éliminé, le reste du réseau apporte immédiatement un effet compensatoire en mettant à jour le taux d'activité individuel des autres neurones sans passer par le biais de la plasticité synaptique. Afin de tester cette théorie, on utilise un modèle développé par Sophie Denève et ses collaborateurs [1] qui proposent que les réseaux de neurone encodent de l'information sensorimotrice efficacement en minimisant le nombre de tir des neurones. Ce modèle a déjà été loué pour avoir des propriétés de robustesse et a une puissance explicative de par la théorie sous-jacente du codage efficace [2]. Dans un premier temps, on veut tester empiriquement cette théorie de compensation immédiate. On utilisera des données expérimentales obtenues chez le zebrafish où au lieu de regarder l'activité isolée des neurones, on regardera l'activité simultanée d'une population de neurones. Avant d'ablater certains neurones chez l'animal, on voudra déterminer les caractéristiques intrinsèques des données quand il est sujet à des efférences sensorimotrices. Ces caractéristiques peuvent être paramétrées en utilisant des méthodes classiques en apprentissage statistique. Des méthodes plus avancées seront aussi développées au cours du projet si nécessaire. Ainsi, en ayant obtenu ces paramètres, on pourra simuler le modèle de réseau de neurones. Après les simulations, on pourra comparer les prédictions du modèle en analysant les résultats des simulations et les données expérimentales. On pourra effectuer l'ablation des neurones dans le modèle et chez l'animal. Le changement d'activité dans le réseau peut alors être quantifié. Si le changement du taux d'activité prédit par le modèle ne correspond pas à ce qui se passe en réalité chez l'animal, on aura alors un moyen de falsifier cette théorie. Dans le cas où la falsification est positive, on passera nos efforts à d'autres approches théoriques afin de pouvoir expliquer ce qui se passe biologiquement. Sinon, ce travail de recherche qui mènera à une thèse doctorale nous aidera à comprendre résolument comment les réseaux de neurones arrivent à maintenir leur robustesse face à des perturbations. [1] Boerlin, Machens and Denève. "Predictive coding of dynamical variables in balanced spiking networks." PLoS Comput Biol9, no. 11 (2013): e1003258. [2] Barrett, Deneve and Machens. "Optimal compensation for neuron death." bioRxiv (2015): 029512

  • Titre traduit

    Robustness of neural networks to external perturbations


  • Résumé

    It has been observed that the brain can maintain its functions even with some neural damage. Patients suffering from Alzheimer's for instance start to experience neuronal death several years before they are clinically diagnosed. This suggests that neural systems are robust to perturbations and our knowledge on how neural networks achieve this is still vague. We suggest that neural networks can automatically compensate for the loss of some of its components without disrupting functionality and we want to investigate in this project, the underlying mechanisms. Our hypothesis is that the compensation can happen on a fast time scale, even before the network is able to update its connections through synaptic plasticity. Instead, in a recurrent network that strives to maintain a tight balance between the level of excitation and inhibition, automatic compensation occurs through an update of the firing rates of the remaining neurons in the population. We build on a theoretical framework of neural networks tuned for efficient coding [1] that display balance between excitation and inhibition and irregular spiking, phenomena that are ubiquitous in the cortex. It was recently shown by Barrett et al. [2] that compensation in these networks is achieved instantaneously without involving synaptic plasticity. They tested the predictions of their model against experimental data where groups of neurons were knocked out pharmacologically. Despite pointing in the right direction, the results are not strong enough as a conclusive test of this theory namely because: (i) using pharmacological knock out procedures usually take 10 to 15 minutes for the effect to take place which is sufficient amount of time for rewiring of the circuit to occur; (ii) observing some neurons shifting in the direction of compensation does not allow us to make full quantitative matches between the model predictions and experimental data – we would actually need population recordings in order the investigate the compensatory effect across the whole network; and lastly (iii) the data collected is limited to understand the compensatory mechanism for all the features that a neuron could be encoding – here, the neurons are recorded in response to single characteristics of the stimulus (for e.g. sensitivity of V1 neurons to the orientations of the stimuli only). In this project, we want to test empirically this theory of immediate compensation in a balanced network. Instead of pharmacological ablation, we shall use data from the larval zebrafish where neural populations are recorded before and after knock-out using optogenetic recording methods [3]. Thus instead of looking at single neurons, we can investigate the compensatory effect across the whole network. This will allow us to measure the changes in activity of the remaining neural components involved in generating a robust population code of sensory inputs after ablation. The recordings are done in the pretectal area and brainstem oculomotor integrator, two areas that have been shown to be encoding sensorimotor variables [4]. Given the spatial and temporal resolution of the available data in these areas, we will be able to measure the changes in the tuning curves of the neurons and compare them with the predictions of the model. Furthermore, the imaging system used in this technique is sufficiently fast to observe the changes in firing rates. This will allow us to measure the immediate adaptation of the remaining neurons upon ablation, thus eliminating the possibility of synaptic plasticity. We want to test the predictions of the network model against the experimental data. To do so, more in depth analysis of this data will be carried out. Prior to ablation, we seek to determine its intrinsic features such as properties of the neurons when decoding the sensorimotor variables that the fish is subject to. Classical techniques such a regression methods can be used to determine the fit of the model to the data. More advanced optimization methods will need to be worked out to improve the accuracy of the fit while preventing overfitting. With satisfactory fitting parameters, we will simulate the neural network before and after knock-out of some neurons. The effect of this ablation can then be quantified using the results of the simulation – the changes in magnitude and direction of the firing rates of the remaining model neurons can be computed. Finally, we will consider the experimental data after ablation. We can check whether using the same intrinsic neuronal features as earlier, we can recover the sensorimotor variables despite the ablation. If so, we can measure the changes in firing rates from the data and compare them with the model predictions established earlier. Otherwise, we can determine the recovery bound of ablation. These results will allow us to check the consistency of this theory, thus providing us with a means to falsify it. If falsified, we will then turn our research endeavours to new frameworks. Otherwise, this work will allow us to elucidate this important question on robustness of neural networks against perturbations. [1] Boerlin, Machens and Denève. "Predictive coding of dynamical variables in balanced spiking networks." PLoS Comput Biol9, no. 11 (2013): e1003258. [2] Barrett, Deneve and Machens. "Optimal compensation for neuron death." bioRxiv (2015): 029512. [3] Orger, Kampff, Severi, Bollmann and Engert. "Control of visually guided behavior by distinct populations of spinal projection neurons." Nature neuroscience 11, no. 3 (2008): 327-333. [4] Portugues, Feierstein, Engert, and Orger. "Whole-brain activity maps reveal stereotyped, distributed networks for visuomotor behavior." Neuron 81, no. 6 (2014): 1328-1343