The physics of Boltzmann machines

Artificial Neural Networks are powerful tools used in the field of artificial intelligence. Inspired by the human brain, they have shown very good results in certain tasks, such as image classification and even in the generation of new images. The problem with neural networks is that we do not understand why they are working so well. To improve the understanding of neural networks we will study a very simple form of them, a so-called restricted Boltzmann machine (RBM). This kind of architecture of neural networks was inspired by statistical physics and therefore we will focus on the understanding of the learning process as a physical process. We will use the language of statistical physics to describe the learning, which is in our understanding nothing else than a process of physical equilibration. After formalizing this process we will focus on the generalization of RBMs to use them for more complex inputs such as for example continuous variables instead of only binary inputs.

Faculty Supervisor:

Peter Wittek

Student:

Partner:

The Institute of Photonic Sciences

Discipline:

Physics

Sector:

Other

University:

University of Toronto

Program: