Thermodynamic machine learning through maximum work production

Abstract Adaptive systems—such as a biological organism gaining survival advantage, an autonomous robot executing functional task, or motor protein transporting intracellular nutrients—must somehow embody relevant regularities and stochasticity in their environments to take full advantage of thermodynamic resources. Analogously, but purely computational realm, machine learning algorithms estimate models capture predictable structure identify irrelevant noise training data. This happens through optimization performance metrics, such model likelihood. If is physically implemented, there sense which estimated are preferred? We introduce the principle that work production most measure for adaptive physical agent compare results maximum-likelihood guides learning. Within class agents efficiently harvest energy from environment, we demonstrate efficient agent’s explicitly determines its architecture how much useful it harvests environment. then show selecting maximum-work given environmental data corresponds finding model. establishes equivalence between nonequilibrium thermodynamics dynamic In this way, maximization emerges organizing underlies systems.