Microchips that learn with local, physics inspired rules

Learning is an emergent property in biological brains: neurons change based on local information to produce global behaviors like memory, adaptation, and recognition. This decentralized design makes brains robust, fault-tolerant, and energy-efficient. Recently, a new framework called physical learning has shown how to achieve the same in circuits, converting global physical phenomena—like the tendency of electrical networks to minimize total power dissipation—into local rules that circuit elements can use to self-adapt and solve problems. Here, we demonstrate that physical learning circuits can be scaled up dramatically in number and down in size by building a microfabricated network of hundreds of low-power learning elements, each just under a hair's width. The entire network can learn tasks using purely local interactions between each element, without a central processor. Inherently robust, this network brings emergent learning into the realm of semiconductor microelectronics, paving the way for learning systems with very large numbers of adaptable degrees of freedom. Likewise, the power consumption for solving tasks on our chip can be as low as one nanowatt—comparable to that of a single biological neuron. This approach holds promise for more power-efficient AI and applications in energy-constrained environments, such as edge electronics and robotics.