First Fairness Benchmark for Spiking Neural Networks Reveals Bias Disparities
A recent study presents the inaugural systematic fairness benchmark for Spiking Neural Networks (SNNs), focusing on three essential aspects: gaps in demographic representation within training datasets, leakage of spurious features (such as using skin color as a proxy for class labels), and inconsistencies in deployment environments on edge devices. This framework combines four cross-demographic datasets with intentional bias injections alongside three neuromorphic hardware simulators (Loihi 2, SpiNNaker). Analysis of 12 cutting-edge SNNs reveals that models trained on biased datasets have a 23% increased rate of false positives for marginalized groups.
Key facts
- First systematic fairness benchmark for SNNs
- Addresses demographic coverage gaps, spurious features, hardware mismatches
- Integrates four cross-demographic datasets
- Uses three neuromorphic hardware simulators: Loihi 2, SpiNNaker
- Evaluates 12 state-of-the-art SNNs
- Biased data leads to 23% higher false positive rates
- Focuses on real-world deployment constraints
- Published on arXiv (2605.27407)
Entities
Institutions
- arXiv