CHORAL what?

"Computational Hardness Of RepresentAtion Learning"

CHORAL aims at developing novel statistical tools to better quantify the performance of neural networks trained on structured data (check this article). This is achieved through a combination of random matrix theory, statistical mechanics and information theory, and a balanced mix of physics techniques together with mathematically rigorous approaches.

Neural nets are amazingly powerful machines, but their theoretical understanding remains limited when compared to their evergrowing applications in science and technology. A related problem of fundamental interest to machine learning that the project will focus on is of “dictionary learning”, a core problem in representation learning.

The CHORAL project and team is generously funded by the European Research Council.