Learning Dynamics

This work presents a fundamental theory of artificial learning that aims to explain existing optimization algorithms used to train artificial intelligence models and to support the development of new ones. The theory is based on a thermodynamic framework, employing geometric and Hamiltonian formalisms to derive the equations governing learning dynamics. These equations provide a physical grounding for the learning process and show that learning can be understood as a thermodynamic process in which models evolve through a succession of equilibrium states, with transitions between them governed by the information the model perceives. article.pdf