Three factor delay learning rules for spiking neural networks

Researchers introduced 'three-factor delay learning rules' for Spiking Neural Networks (SNNs), enabling online learning of synaptic and axonal delays. This approach improves classification accuracy by up to 20% over weight-only baselines. On the SHD speech recognition dataset, it matches offline methods while reducing model size by 6.6x and cutting inference latency by 67%, with only a 2.4% accuracy drop. The method is designed for real-time operation in resource-constrained neuromorphic processors.