Distributed neuromorphic computing with temporal co-optimization of spiking dynamics

Neuromorphic processors and event-based sensors enable always-on perception, yet training their computational building blocks, namely spiking neural networks (SNNs) across heterogeneous edge devices, is constrained by bandwidth, latency, and energy. This paper studies the systems problem of distributing continuous-time SNN training when computation and communication are driven by sparse spike events. Building on a previously developed exact event-driven single-device learner (ExactTrain) as a local training primitive, we develop distributed protocols, an AER communication abstraction, and the associated convergence, traffic, and system analyses. In particular, we formulate: (i) FedSNN, a federated protocol that matches centralized exact-gradient training under full participation and a one-full-local-gradient regime, (ii) SplitSNN, a split protocol that partitions a model across devices and exchanges address-event representation (AER) spike and error packets, and (iii) HybridSNN, a two-tier split/federated hierarchy that exposes an explicit accuracy-communication trade-off.Experiments on the N-MNIST and DVS Gesture benchmarks evaluate centralized equivalence, split and hybrid traffic, matched-configuration and byte-accounted baseline comparisons, scalability, sensitivity to non-IID data and partition location, and trace-driven network emulation with injected bandwidth limits, jitter, and packet loss. The results reveal that the proposed protocols preserve the timing benefits of exact event-driven local learning while replacing dense activation exchange with sparse spike/error traffic whose scaling is predictable from event rates and cut connectivity.