Kyungeun Kim, Amanuel Anteneh, Israel Klich, Olivier Pfister, J. M. Schwarz
PCPL is a general framework for learning by contrasting perturbation responses in physical systems, enabling learning from the system's own response without external backpropagation.
Existing physical learning methods (e.g., Equilibrium Propagation) are limited to specific energy-based systems or require centralized gradient computation by an external processor. A more general and autonomous physical learning system is needed.
Leverage measurable contrasts between physical states produced by controlled changes to inputs, boundary conditions, parameters, or interpreter functions. PCPL unifies Equilibrium Propagation and Frequency Propagation, using contrast-based updates that reflect either local sensitivities or global inverse-problem structure. Learning geometry emerges implicitly from the system's physical response without an external processor.
Successfully learned classification tasks on two platforms: spring networks and continuous-variable photonic circuits. The photonic circuit was trained to implement analog multiplication, demonstrating progress toward more autonomous physical learning systems.