AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
The Limits of Deep Learning
Janik: I'm a huge fan of Francois as well and he often talks about the limitations of deep learning in the sense that it's they're very good for interpolation given a sufficiently dense sampling of the input space. What are the limitations of using deep learning for this kind of work Guillaume Lamppole had this very nice paper on the on this automated integration which was eye-opening to Janik. He says auto formalization is much more important than self-supervised training or exploration of mathematics becauseAuto formalization gives us a grounding so we cannot open and easily go and discover everything but there are multiple problems with some random direction mathematics just too big if you don't have