Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Steven Motta, Gioele Nanni

Abstract

Federated learning on neuromorphic hardware remains unexplored because on-chip spike-timing-dependent plasticity (STDP) produces binary weight updates rather than the floating-point gradients assumed by standard algorithms. We build a two-node federated system with BrainChip Akida AKD1000 processors and run approximately 1,580 experimental trials across seven analysis phases. Of four weight-exchange strategies tested, neuron-level concatenation (FedUnion) consistently preserves accuracy while element-wise weight averaging (FedAvg) destroys it (p = 0.002). Domain-adaptive fine-tuning of the upstream feature extractor accounts for most of the accuracy gains, confirming feature quality as the dominant factor. Scaling feature dimensionality from 64 to 256 yields 77.0% best-strategy federated accuracy (n=30, p < 0.001). Two independent asymmetries (wider features help federation more than individual learning, while binarization hurts federation more) point to a shared prototype complementarity mechanism: cross-node transfer scales with the distinctiveness of neuron prototypes.

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

Abstract

Federated learning on neuromorphic hardware remains unexplored because on-chip spike-timing-dependent plasticity (STDP) produces binary weight updates rather than the floating-point gradients assumed by standard algorithms. We build a two-node federated system with BrainChip Akida AKD1000 processors and run approximately 1,580 experimental trials across seven analysis phases. Of four weight-exchange strategies tested, neuron-level concatenation (FedUnion) consistently preserves accuracy while element-wise weight averaging (FedAvg) destroys it (p = 0.002). Domain-adaptive fine-tuning of the upstream feature extractor accounts for most of the accuracy gains, confirming feature quality as the dominant factor. Scaling feature dimensionality from 64 to 256 yields 77.0% best-strategy federated accuracy (n=30, p < 0.001). Two independent asymmetries (wider features help federation more than individual learning, while binarization hurts federation more) point to a shared prototype complementarity mechanism: cross-node transfer scales with the distinctiveness of neuron prototypes.
Paper Structure (58 sections, 1 equation, 2 figures, 14 tables)

This paper contains 58 sections, 1 equation, 2 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Neuromorphic Computing and SNNs.
  4. Federated Learning.
  5. Federated Learning on Neuromorphic Hardware.
  6. Few-Shot Learning and Transfer Learning.
  7. Proposed Method
  8. Two-Stage Pipeline: Bridging IP Versions
  9. Wide feature variants.
  10. Federation Strategies
  11. FedAvg.
  12. FedUnion.
  13. FedBest.
  14. FedMajority.
  15. Reporting convention.
  16. ...and 43 more sections

Figures (2)

  • Figure 1: Two-stage pipeline. Audio is converted to MFCC spectrograms, processed by the DS-CNN feature extractor (software, v2) to produce int8 features, optionally projected to 128/256 dimensions, binarized to 1-bit, then classified by the STDP edge learner running on-chip on the Akida AKD1000 (hardware, v1).
  • Figure 2: System topology. Two RPi 5 nodes with Akida AKD1000 coprocessors exchange STDP-learned weight vectors (20--40 KB) over a direct Ethernet link. Each node runs the two-stage pipeline (feature extraction followed by on-chip STDP learning). Four federation strategies merge the weight matrices. A Mac provides remote monitoring via Tailscale.