Physical Society Colloquium

Biophysics of decision-making and machine learning

Paul François

Department of Physics
McGill University

Many biological phenomena are best understood with the help of out of equilibrium statistical mechanics. A classical example is DNA replication which can “beat” thermodynamics using kinetic proofreading. This “wrong” vs “correct” pairing is one of the simplest instance of cellular decision-making.

In the context of immune discrimination (“self” vs “non self”), we have developed and validated experimentally a more general theory that we named adaptive kinetic proofreading. However, this family of models presents new unexpected sensitivity to specific perturbations, which explains how some mutating agents (e.g. HIV virus) can escape the immune system. This is actually reminiscent of a general problem in machine learning, namely the existence of so-called adversarial examples. For instance, few changes in carefully chosen background pixels barely visible by eye can strongly perturb image classification by neural networks. In this Colloquium, I will introduce common concepts related to decision-making in both the biophysical and machine learning context. Our work suggests a (simple) connection between the two problems, that I will describe.