Did Richard Feynman help seed a key machine learning technique in the 60s?

I struggled to learn machine learning. I was used to variational tricks, MCMC samplers, and discreet Taylor expansions from years of physics training. Now the concepts were mixed up. The intuitive models of physical systems were replaced by abstract models of ‘data’ and amechanical patterns of cause and effect.

I had to fit these fields together. Physics and machine learning are intricately connected, but it is taking me years to make the overlaps precise. This process requires representing the new with the familiar, mapping jargon from one field to another.

A simple model of magnets—the Ising model—will help illustrate the rich connection between these fields. We first analyze this model with physics intuition. Then we derive the variational principle in physics and show that it recovers the same solution.

We then discover how that very same variational principle in physics opens a window into machine learning. We identify Boltzmann distributions as exponential families to make the mapping transparent, and show how approximate posterior inference is scaled to massive data thanks to the variational principle.

If you have a physics background, I hope you will have a better sense of machine learning and be able to read papers in the field. If you are a machine learner, I hope you will have the context to read a statistical physics paper about mean-field theory and the Ising model.

*If this article is confusing, falls short of these goals, or could be improved in any way please email me, @ me, or submit a pull request.*