Building a Physical Learning Network

Computational networks are flexible tools capable of a vast array of tasks, from computer vision to motor control. However, unlike biological networks like the brain, they utilize non-local learning rules (e.g. minimizing a global cost function) and thus require external computation. Here we build a physical system - an electrical network of variable resistors – capable of learning a range of tasks: a physical learning network. This system evolves according to simple local rules, allowing it to learn using energy minimization (in the form of Kirchoff’s Laws) to naturally execute the required ‘computation’. This type of network may prove useful for a variety of purposes like flexible sensors or controllers in situations where an entire CPU is untenable, and will allow us to explore the building blocks of learning in a fully understood system.