Vous êtes ici : Accueil > Entités de recherche > NeuroSpin > The circuits and dynamics underlying generative processes in neural networks


Conférence NeuroSpin

The circuits and dynamics underlying generative processes in neural networks

Du 07/10/2019 au 07/10/2019
NeuroSpin, Amphithéâtre

Michel  Besserve from Max Planck Institute for Intelligent Systems will give a talk at NeuroSpin on October 7th. 

He's interested in understanding structure and dynamics of coordination in neural networks. 

Short abstract:

Our subjective experience as we dream, remember and even perceive the world provides countless illustrations of a simple fact: the brain actively generates and manipulates representations. Although Neuroscience has made considerable progress in uncovering the neural substrate of this ability at multiple levels, the precise network mechanisms allowing these representations to reliably emerge and evolve with our environment are still largely elusive.

First, we study two phasic events supporting these functions in the mammalian brain: hippocampal ripples associated with episodic memory replay and ponto-geniculo-occipital waves associated with dreams. Non-human primate multi-modal data combined with computational modelling sheds light on two crucial elements controlling their dynamics: excitatory-inhibitory micro-circuits impose a tight control on the sequential content of the elicited representations, while neuromodulatory centers located in the brainstem coordinate the emergence of these representations in a state dependent way. Finally, we find evidence that these brain-wide transient activities leverage long-term potentiation and depression mechanisms to foster precise network reconfigurations during sleep.

In a second part, we investigate the modularity of generative representations, which are likely key to generalization abilities of autonomous agents. We study this in deep generative models, a class of artificial neural networks that can generate complex data such as realistic images, and develop a causal framework based on the principle of independence of mechanisms to address the modularity of these networks. We show such modular organization can be uncovered by intervening on groups of neurons, allowing to generate meaningful “counterfactual images” that replace targeted properties of the original image, such as elements of the background.

Overall, the join study of artificial and biological generative processes helps us elucidate the key principles and associated mechanisms required to maintain efficient and adaptive distributed representation systems in neural networks.


Infos Pratiques


11.00 - 12.00

Haut de page


Flux rss