Information bath and learning processes along geodesics

Author:

Arnab Barua, Haralampos Hatzikirou, Sumiyoshi Abe

Keyword:

Condensed Matter, Disordered Systems and Neural Networks, Disordered Systems and Neural Networks (cond-mat.dis-nn), Biological Physics (physics.bio-ph)

journal:

date:

2024-01-09 00:00:00

Abstract

Learning is a fundamental characteristic of living systems, enabling them to comprehend their environments and make informed decisions. These decision-making processes are inherently influenced by available information about their surroundings and specific objectives. An intriguing perspective is that each process could be optimal under a given set of conditions. Here, a geometric method is introduced for analytically exploring such a perspective. The probability distribution describing the state of the composite system of the environment, termed the information bath, and a decision-maker is discussed on the basis of the entropic concept. This enables one to study the system in analogy with thermodynamics. Learning processes are expressed as the changes of parameters contained in the distribution. For a geometric interpretation of the processes, the manifold endowed with the Fisher-Rao metric as the Riemannian metric is considered. This framework allows one to conceptualize the optimality of each process as a state change along a geodesic curve on the manifold, which is referred to here as geodesic learning. Then, the bivariate Gaussian model is presented for illustrating this approach, and its rigorous solution is obtained. There, the correspondence between learning and cooling is made manifest. Evolution of the entropy is evaluated for geodesic learning. The findings are then interpreted in the context of adaptation. Geodesic learning plays a role of the case of reference and may give novel insights into the mechanics of learning in living systems.

PDF: Information bath and learning processes along geodesics.pdf