Fathom the transition in a learning algorithm through irradiation-tuned pore condensation

Abstract Phase transitions represent fundamentally interdisciplinary phenomena occurring across a wide range of sizes and time horizons beginning at the subatomic scale. They also appear in formal mathematical problems including optimisation, inference or learning tasks. In any case, critical points are significant for marking where a system’s behaviour decisively alters. Conveniently, the characteristics of state changes allow revealing shared collective mechanisms from seemingly unrelated microscopic details. Very recently, for a popular learning ansatz a novel transformation-like effect has been discovered. In general, its information processing aims to iteratively split provided numerical inputs. In particular, the structural growth of the resulting partitions shows striking parallels with molecular adsorption in confined geometries and the accompanying vapour–liquid conversion. Accordingly, statistical mechanics and chemical modelling are employed in this paper to provide an initial high-level understanding from a comparative perspective. Specifically: (i) previous complex empirical findings are reduced to a minimal viable example comprising a simplified algorithm and an expressive one-dimensional dataset; (ii) essential computational parameters are systematically probed; (iii) those parameters are heuristically mapped onto thermodynamic quantities to explore analogies of learning with condensation in porous media; and (iv) kinetic Monte Carlo simulations are performed to flexibly illustrate microscopic chemical details. The study outcomes indicate that the two key attributes facilitating the structural rearrangement during learning have an intuitive role: they correspond to pore diameter as well as the fluid–fluid interparticle cohesion. Experimentally both could be modulated by focused irradiation.