Maths and Physics of Learning in Artificial and Biological Networks (MaPLe)

This Nordita programme asks what learning is and why both artificial and biological neural networks learn so efficiently, and how features of the training data shape learning. Despite the success of deep networks, progress in deep learning has been largely heuristic - arguably comparable to thermodynamics before Carnot. Over the past decade, promising ideas have emerged from statistical mechanics and from algebraic topology, while detailed connectivity and activity data from biological networks now allow researchers to extract signatures of learning in animals. The programme brings together leading physicists, mathematicians and neuroscientists to synthesise these theoretical and experimental advances towards a theory of learning, better mathematical tools for training artificial networks efficiently, and sharper experiments and hypotheses about the neural correlates of learning. Confirmed invited participants include John Hertz, Christopher Lynn, Massimiliano Mella, Tomas Pajdla, Prasad Perlekar, Mason Porter, Stefano Sarao Mannelli, Vahid Shahverdi, Haim Sompolinsky and Rene Vidal.