Federated Few-Shot Learning on Neuromorphic Hardware

Code for the paper: Federated Few-Shot Learning on Neuromorphic Hardware: An Empirical Study Across Physical Edge Nodes (Motta & Nanni, 2026).

This repository contains the full experimental framework for federated few-shot class learning using BrainChip Akida AKD1000 neuromorphic processors with on-chip STDP. Two physical Raspberry Pi 5 nodes, each equipped with an Akida PCIe accelerator, learn novel keyword classes from limited samples and exchange neuron-level weight maps over a direct Ethernet link.

Hardware Requirements

• 2x Raspberry Pi 5 (8 GB recommended)
• 2x BrainChip Akida AKD1000 (PCIe HAT)
• Direct Ethernet link between Pis (10.0.0.x subnet)
• Orchestrator machine (Mac or Linux) with SSH access to both Pis

Software Requirements

On each Raspberry Pi:

• Python 3.12 with virtualenv at ~/akida-env/
• akida, akida-models, cnn2snn (BrainChip SDK)
• numpy, scipy, scikit-learn

On the orchestrator:

• Python 3.10+
• numpy, scipy, scikit-learn, paramiko

Repository Structure

File	Description
config.py	All hyperparameters, network topology, and experimental grid definitions
run_experiment.py	Main entry point for running experiments
orchestrator.py	Coordinates multi-node experiments from the orchestrator machine
node_worker.py	Worker process running on each Raspberry Pi
federation.py	Federation strategies: FedUnion, FedAvg, FedBest, FedMajority, FedSelective
finetune_dscnn.py	Feature extractor fine-tuning (DS-CNN backbone) with wide feature support
data_loader.py	Google Speech Commands v0.02 dataset loading and MFCC extraction
baselines.py	Software baselines (k-NN, linear, MLP) for comparison
baselines_pi.py	On-device software baselines running on Raspberry Pi
comprehensive_sweep.py	Full Phase A-F experimental sweep (binarization, disjoint, wide features, multi-round)
hyperparam_sweep.py	STDP hyperparameter grid search
analyze_results.py	Statistical analysis, bootstrap CIs, Cohen's d, result tables
setup_claudio.sh	Pi setup script
run_autonomous.sh	Autonomous experiment runner
run_autonomous_v2.sh	Updated autonomous runner with comprehensive sweep

Quick Start

1. Prepare Data

Download Google Speech Commands v0.02 and extract MFCC features:

python -m run_experiment --data-only

2. Deploy to Pis

Ensure SSH keys are configured for both Pis, then:

python -m run_experiment --num-trials 1 --num-rounds 1

This deploys code and data to both Pis, runs a single trial, and collects results.

3. Full Experiment

To reproduce the paper results (10 trials, 5 federation rounds, all baselines):

python -m run_experiment \
    --num-trials 10 \
    --num-rounds 5 \
    --shared-thresholds \
    --run-baselines

4. Comprehensive Sweep (Phases A-F)

To run the full experimental sweep including binarization comparison, disjoint validation, wide features, and multi-round analysis:

python comprehensive_sweep.py

5. Analysis

Generate statistical analysis from collected results:

python analyze_results.py

Federation Strategies

• FedUnion: Concatenate neuron populations from all nodes (neuron-level aggregation)
• FedAvg: Average weight vectors across matching neurons (weight-level aggregation)
• FedBest: Select the single best-performing node's model
• FedMajority: Ensemble majority vote across node predictions
• FedSelective: Selective neuron merging based on activation thresholds

Experimental Phases

The comprehensive sweep (comprehensive_sweep.py) runs six phases:

Phase	Description	Runs
A	Baseline (pre-fine-tuning)	--
B	42-config hyperparameter sweep (7 nw x 3 npc x 2 lc)	420
C	Binarization comparison (mean/median/entropy) on top configs	150
D	Disjoint class validation (extractor never sees target classes)	50
E	Wide feature scaling (128-dim and 256-dim)	130
F	Multi-round federation stability analysis	60

Total: approximately 800 experimental runs on physical hardware.

Citation

If you use this code, please cite:

@article{motta2026federated,
  title={Federated Few-Shot Learning on Neuromorphic Hardware: An Empirical Study Across Physical Edge Nodes},
  author={Motta, Steven and Nanni, Gioele},
  year={2026}
}

License

MIT License. See LICENSE for details.