Rectified Linear Postsynaptic Potential Function for Backpropagation in Deep Spiking Neural Networks.

Area of Science:

• Neuromorphic computing
• Artificial intelligence
• Computational neuroscience

Background:

• Spiking neural networks (SNNs) mimic biological neurons for efficient, event-driven processing.
• Training deep SNNs (DeepSNNs) is challenging as standard backpropagation (BP) is not directly applicable.
• Temporal coding in SNNs uses spike timing for information representation.

Purpose of the Study:

• To investigate the limitations of error BP in DeepSNNs.
• To propose a novel learning algorithm for DeepSNNs that utilizes spike timing.
• To demonstrate ultralow-power inference using DeepSNNs on neuromorphic hardware.

Main Methods:

• Developed a rectified linear postsynaptic potential (ReL-PSP) function for spiking neurons.
• Introduced a spike-timing-dependent BP (STDBP) learning algorithm for event-driven training.
• Implemented and evaluated DeepSNNs on a neuromorphic inference accelerator.

Main Results:

• DeepSNNs trained with STDBP achieved state-of-the-art classification accuracy on the MNIST dataset.
• Neuromorphic hardware demonstrated ultralow power consumption (0.751 mW) and low latency (47.71 ms).
• Spike timing dynamics were shown to be crucial for information encoding, synaptic plasticity, and decision-making.

Conclusions:

• The proposed STDBP algorithm effectively trains DeepSNNs, enabling high accuracy and efficient inference.
• Spike timing is a key factor for advancing SNNs and neuromorphic hardware design.
• This research offers a new perspective for developing future DeepSNNs and neuromorphic systems.