A brain-inspired electronic learning machine

Contrastive learning algorithms have recently been proposed for training physical networks such as mechanical, flow and electrical networks to perform arbitrarily complex machine learning tasks not by minimizing a global cost function as in artificial neural networks, but in a manner more similar to the brain, using only local information. To date, however, they have only been implemented {it in silico} due to the need for a central processor and memory storage in order to compare the response of the network to two different sets of boundary conditions and accordingly update the network elements. Here, we introduce a method to implement a physics-driven contrastive learning scheme in the laboratory for a network of variable resistors, using circuitry to compare the response of {it two} networks that have identical resistances but are subjected to the two different sets of boundary conditions. With this innovation, we demonstrate how our system optimizes its resistances and effectively trains itself, without use of a central processor or information storage, to perform specified allostery, regression, and classification tasks. Once the system is trained, the desired tasks are subsequently performed rapidly and automatically by the physical imperative of local currents to adjust in order to minimize power dissipation for the given voltage inputs. Our twin-network laboratory approach may be readily scaled to extremely large or nonlinear networks using modern microfabrication techniques. Moreover, our implementation has an enormous scaling advantage compared to {it in silico} implementations because the forward computation is done by the physics and without need for information storage; even a modestly larger laboratory network of 500 nodes will outperform its in silico counterpart. Finally, we demonstrate that such learning systems are robust to extreme damage due to their decentralized character.

Speaker: Douglas Durian, University of Pennsylvania