A Thermodynamic Perspective on Machine Learning

Modern machine learning methods increasingly face questions of efficiency, robustness, and the fundamental limits of computational processes. This talk will present an approach that transfers concepts and intuitions from classical thermodynamics into the realm of machine learning. Applications of this perspective will be demonstrated in two domains: topic modeling and neural network pruning.

The main topics of the talk include:

— analogies between thermodynamic systems and machine learning models;

— entropy and free energy in the context of information and uncertainty in machine learning;

— topic modeling and the problem of selecting the number of clusters;

— neural network pruning and the problem of choosing an appropriate sparsity level.

The target audience includes machine learning researchers, AI practitioners, physicists interested in interdisciplinary approaches, and developers seeking a deeper understanding of the fundamental principles underlying modern learning systems.