Event-based backpropagation can compute exact gradients for spiking neural networks

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Deep learning advances rely on backpropagation for non-spiking neural networks.
• Training spiking neural networks (SNNs) with backpropagation is challenging due to discrete spike events.

Purpose of the Study:

• To derive an exact backpropagation algorithm for continuous-time spiking neural networks.
• To enable gradient-based learning in SNNs without approximations.

Main Methods:

• Applied the adjoint method with partial derivative jumps to derive the backpropagation algorithm.
• Developed EventProp for event-based, temporally and spatially sparse gradient computation.
• Utilized EventProp to train SNNs on Yin-Yang and MNIST datasets.

Main Results:

• Achieved competitive performance in SNN training using EventProp.
• Demonstrated backpropagation through discrete spike events without approximations.
• Enabled gradient computation at spike times for exactness.

Conclusions:

• EventProp facilitates rigorous study of gradient-based learning in SNNs.
• Provides insights for implementing SNNs in brain-inspired hardware.
• Advances the training of biologically plausible neural networks.