In this work, we decided to look at the relations between two seemingly rather disconnected concepts - free energy extraction, and machine learning. it turns out though, that the two concepts hold rich relations. Notably, for a demon to be able to extra the greatest work from a pattern, he needs to have a mental model of the pattern. Thus, we find that the amount of free energy the demon can extract coincides with a popular measure of model efficacy in machine learning (the maximum likelihood). This then allows us to think about training models from the physical perspective of a demon learning the best method to extract work.
- Thermodynamic machine learning through maximum work production
A Boyd, J Crutchfield, M Gu
New Journal of Physics
