Deep learning is having a big impact in many domains of science and technology. While all these developments are exciting and fascinating, and were inconceivable just a few years ago, they have not led to a significantly deeper understanding of the principles underlying these application domains. A main reason for this is the lack of understanding of the principles behind deep learning itself. Theoretical approaches rooted in statistical physics provide important insights to this important question. This program is aimed at bringing together researchers with a particular emphasis on cognitive science and neuroscience.

This focus arises from the special relationship between deep learning and neuroscience. For other domains, deep learning acts more like a parametric model, where a parallel between the mechanistic aspects of deep learning and the phenomenon modeled is very hard to draw. In contrast, neural networks provide mechanistic models of how the brain functions. Indeed, deep learning models of the brain can be predictive of neuronal activity at the single cell level and provide mechanistic insight into computations in the brain. Therefore, advances in deep learning theory may shed light onto how the brain works. However, for this to happen, new theories need to demonstrate how cognitive capabilities such as planning, reasoning and memory can emerge from neural dynamics while addressing the role of biological features–including cell types, biologically plausible learning rules, and biological network architectures–in cognition. Our program will bring together researchers in this domain to identify opportunities for progress.