Quantum Deep Learning

Daniel Persson, a mathematical physicist at Chalmers, presented a research project in its first year: using ideas from physics to understand why deep learning works and generalises as well as it does, rather than to build new algorithms. The project is one of several that the Wallenberg AI, Autonomous Systems and Software Program (WASP) had started at the mathematics department.

He used restricted Boltzmann machines, whose energy function comes from the Ising model, as a test case, and discussed the proposed analogy between training stacked networks and the renormalisation-group flow of physics, in which a system is coarse-grained step by step. He then turned to the other direction, from machine learning as a tool for quantum many-body systems to entanglement entropy and its area law as a possible guide for network architectures, with a parallel to the entropy of black holes.

In the questions, he agreed that the renormalisation analogy holds for some networks only, and that how far it extends is an open question.