Training Deep Spiking Neural Networks Using Backpropagation.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Deep spiking neural networks (SNNs) offer potential for low-latency, energy-efficient computation via event-based processing.
• Training SNNs is challenging due to the non-differentiable nature of spike events.

Purpose of the Study:

• Introduce a novel, direct training method for deep SNNs.
• Enable error backpropagation directly on spike signals and membrane potentials.
• Improve accuracy and efficiency compared to existing SNN training techniques.

Main Methods:

• Treating membrane potentials as differentiable signals, with spike discontinuities modeled as noise.
• Implementing an error backpropagation mechanism analogous to conventional deep networks.
• Evaluating the framework on MNIST and N-MNIST benchmarks.

Main Results:

• Achieved a >3x reduction in error rate on N-MNIST compared to prior SNNs.
• Outperformed conventional convolutional neural networks (CNNs) on N-MNIST.
• Demonstrated equivalent accuracy to CNNs on MNIST with significantly fewer operations on N-MNIST.

Conclusions:

• The proposed differentiable approach enables effective training of deep SNNs.
• Deep SNNs trained with this method offer competitive accuracy and superior computational efficiency.
• This technique advances the practical application of SNNs in areas requiring low power and fast processing.