| Abstract: | Though no match to practical breakthroughs, our theoretical grasp of deep learning has steadily widened during the last decade. This talk will review the past misconceptions and recent new conceptions that have formed, taking a physics standpoint. In particular, I will describe some concrete links to entropy, order by disorder, field theory, and renormalization. I will also discuss the paradoxical nature of trying to understand models meant to solve problems we cannot solve analytically, and ways around this. |
