On physical processes that work like learning algorithms

Understanding the physical origins of learning in natural systems requires establishing a correspondence between physical processes and learning algorithms. Typically, learning algorithms comprise two key phases: signaling and weight updating. Inspired by Physarum polycephalum, we introduce a novel signaling mechanism for optimizing flow networks. In our approach, error information is encoded into advective chemical signals, allowing individual network components to collectively minimize a global cost function.

Building on this concept, we propose the Multi-mechanism Learning (MML) framework, in which input and feedback signals are encoded into two distinct, non-interfering physical quantities. Within this framework, we develop a backpropagation-like learning process that uses only local information to perform gradient descent on a global cost function.

Furthermore, we posit that glassy dynamics—characterized by slow, history-dependent behavior—may represent nature’s strategy for implementing weight updates. Leveraging this concept, we demonstrate how a granular system self-organizes in response to external driving and how this mechanism can be used to exhibit allostery-like behavior. This training paradigm eliminates the need for explicit weight update rules, enabling the system to self-organize naturally toward targeted outcomes.

Our findings offer new insights into how physical processes can emulate learning, with potential implications for understanding adaptation and self-organization in complex biological systems.