Going deep, reaching further: What can Boltzmann machines efficiently represent?

While universal approximation theorems theoretically underpin the representation capabilities of machine learning models, they have two shortcomings: They do not focus on efficient representations and they tend to be insensitive to depth and other heuristically important properties.

By Tobias Reinhart, Gemma De les Coves · 1-minute read