New SNN fairness benchmark reveals 41% accuracy gap on edge devices

A team of researchers from multiple institutions has unveiled the first comprehensive fairness benchmark for Spiking Neural Networks (SNNs), addressing critical gaps in evaluating bias and hardware-driven disparities. The work, led by Hudi He and published on arXiv, systematically tests how SNNs perform across diverse demographic groups when subjected to three real-world challenges: biased training data, spurious feature leakage (e.g., skin tone as a proxy for labels), and deployment-environment mismatches. The benchmark integrates four cross-demographic datasets with controlled bias injections and three neuromorphic hardware simulators—Intel's Loihi 2 and SpiNNaker—to assess isolated fairness-performance trade-offs under resource constraints.

Standardized evaluations of 12 state-of-the-art SNNs yielded stark results: models trained on biased data exhibited 23% higher false positive rates for underrepresented groups, while hardware limitations—such as reduced spike precision on edge devices—amplified accuracy gaps by up to 41%. Critically, bias mitigation strategies that work well in cloud-based SNNs often degrade when deployed on resource-limited neuromorphic chips, highlighting the need for co-design principles that jointly optimize fairness and hardware efficiency. The benchmark is open-source, aiming to bridge algorithmic fairness research with neuromorphic engineering for socially critical applications like healthcare and autonomous systems.