Federated reservoir intelligence for distributed temporal and neuromorphic learning

Abstract Federated learning lets many devices train a shared model without pooling their private data, yet applying it to time-varying and event-driven signals is expensive: deep recurrent and spiking networks must exchange large numbers of parameters in every communication round. Here we show that combining federated learning with reservoir computing reduces this bottleneck. Each device runs a fixed, untrained recurrent network and shares only a compact linear output layer or its summary statistics. With shared reservoir features and full participation, sharing statistics reconstructs the centralized solution in one round; the iterative alternative converges quickly under intermittent participation; personalization and exponentially-discounted temporal updates also follow in closed form. On brain–computer interface and event-camera benchmarks, the approach matches locally trained models or deep baselines in accuracy-matched settings while reducing communication severalfold, and a compressed variant lowers communication by more than two orders of magnitude where an accuracy trade-off is acceptable. It requires no backpropagation, suiting low-power neuromorphic hardware.