Human-Interpretable Concept Learning via Information Lattices

Is it possible to learn the laws of music theory directly from raw sheet music in the same human-interpretable form as a music theory textbook? How little prior knowledge needs to be encoded to do so? We consider these and similar questions in other topical domains, in developing a general framework for automatic concept learning. The basic idea is an iterative discovery algorithm that has a student-teacher architecture and that operates on a generalization of Shannon’s information lattice, which itself encodes a hierarchy of abstractions and is algorithmically constructed from group-theoretic foundations. In particular, learning this hierarchy of invariant concepts involves iterative optimization of Bayesian surprise and entropy. This gives a first step towards a principled and cognitive way of automatic concept learning and knowledge discovery. We further discuss applications in computational creativity, AI safety, and AI ethics.

Video: https://vimeo.com/332021301