Alexis Dubreuil
Institut de la Vision, Sorbonne Université
http://alexisdubreuil.fr/research-2
Mechanics of neural computation
Describing how neural networks perform computations requires to understand how a macroscopic dynamics can emerge from a specific network structure. Statistical physics approaches to neural networks are particularly well suited to provide such a description. I will start by introducing the Hopfield model for associative memory, which has been extensively studied in the 80’s from a statistical physics point of view, and then show how this type of approach allows to address modern questions in neural computation. But as these standard approaches rely on the ability of modelers to propose neural structures implementing the desired computations, they quickly becomes difficult as one considers increasingly complex computations. I will present an approach, that combines deep-learning algorithms and neural network theory, to leverage this problem. This will be the occasion to further illustrate how the macroscopic dynamics of well defined order parameters relates to computation.