Reproduce “Three factor delay learning rules for spiking neural networks” on OpenResearch

About this paper

Luke Vassallo, Nima Taherinejad

Three-factor delay learning extends online training in spiking neural networks beyond adjusting synaptic weights, allowing synaptic and axonal timing parameters to adapt from local eligibility traces and task-level error signals. On spoken-digit recognition, learned delays raise accuracy from 79.46% to 92.64% in a 128-neuron feedforward SNN, while sparse delay-based models achieve comparable performance with fewer parameters and lower estimated hardware cost.

Key insights

• Random or fixed delays already provide much of the accuracy gain, but learning the delays adds further improvements, especially in small and sparse networks.
• Axonal delays are more parameter- and memory-efficient because one delay serves all outgoing connections, whereas synaptic delays offer finer temporal flexibility and generally higher absolute accuracy.
• Delay learning can make compact models competitive with larger weight-only networks, but the Gaussian delay-gradient approximation is non-causal and introduces buffering overhead for hardware learning.

Summary by alphaXiv.