Learning with artificial and natural neural networks: trade-offs in energy consumption and representations

Laboratory of Physics of the École Normale Supérieure, PSL Research and CNRS, Paris, France

Abstract

Decades after Hopfield and Hinton’s seminal works on neural computation and Boltzmann machines, the use of neural networks in machine learning has revolutionized artificial intelligence. Physics, with the help of neuroscience, still has a lot to say on many issues. We here discuss two of them, energy consumption and representations.