Understanding Deep Neural Networks with Game TheoryUNDERNEATH (JCJC ANR Project) - coordinator : Ievgen REDKO

The theoretical analysis of deep neural networks (DNN) is arguably among the most challenging research directions in machine learning right now, as it requires scientists to lay novel statistical learning foundations to explain their behavior in practice. In this proposal, we aim to explore the interplay between DNNs and game theory by considering the widely studied class of congestion games with the goal of relating them to both linear and non-linear DNNs and to the properties of their loss surface. Beyond retrieving the state-of-the-art results from the literature in a principally new way, we expect that our proposal will provide a very promising novel tool for analyzing DNNs and will allow solving concrete open problems related to 1) characterizing the DNNs optimization inefficiency depending on the algorithmic choices, such as their architecture, activation, and loss function used and 2) proposing new optimization strategies with strong convergence guarantees.