Federated Few-Shot Learning on Neuromorphic Hardware: An Empirical Study Across Physical Edge Nodes

Researchers Steven Motta and Gioele Nanni have conducted the first large-scale empirical study of federated learning on physical neuromorphic hardware, a significant step toward efficient, privacy-preserving AI at the edge. They built a two-node federated system using BrainChip's Akida AKD1000 processors—specialized chips that mimic the brain's spiking neural networks—and ran approximately 1,580 experimental trials. The core challenge was that neuromorphic chips use spike-timing-dependent plasticity (STDP) for on-chip learning, which produces binary weight updates, unlike the floating-point gradients required by standard federated algorithms like FedAvg.

Their key finding was that the standard federated averaging (FedAvg) method, which averages model weights, completely destroyed accuracy on this hardware. In contrast, a novel strategy they call 'FedUnion,' which concatenates neuron-level prototypes across nodes, consistently preserved it, with the performance difference being statistically significant (p = 0.002). The study also revealed that scaling the feature dimensionality from 64 to 256 was critical, yielding a best-strategy federated accuracy of 77.0%. The results point to a 'prototype complementarity' mechanism, where successful cross-node knowledge transfer depends on the distinctiveness of the neuron prototypes learned by each device.