Equilibrium Propagation

An interactive demonstration of equilibrium propagation training a tiny network with one clamped input, two hidden neurons, and one output. Each training step runs a free relaxation and a weakly-clamped nudged relaxation toward a target, then updates each weight by a local contrastive Hebbian rule. The network diagram shows weights as colored edges and free-phase activations as node fills; small deltas show how the target nudge shifts each hidden neuron. A loss sparkline falls as the free-phase output converges to the target.