Spike-Based Approximate Backpropagation Algorithm of Brain-Inspired Deep SNN for Sonar Target Classification.

Area of Science:

• Neuromorphic Computing
• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking neural networks (SNNs) offer ultralow energy consumption and efficient spatiotemporal processing, inspired by the brain.
• Training deep SNNs directly is challenging due to the discontinuous nature of spiking neuronal activation functions.
• Current SNNs often lag behind artificial neural networks (ANNs) in performance.

Purpose of the Study:

• To propose a novel method for the direct, end-to-end training of deep brain-inspired SNNs.
• To enable SNNs to achieve performance comparable to ANNs while retaining their energy efficiency advantages.

Main Methods:

• Introduction of the spike-based approximate backpropagation (SABP) algorithm.
• Development of a general brain-inspired SNN framework.
• Integration of SABP with the SNN framework for direct deep SNN training.

Main Results:

• The proposed method achieves classification accuracy close to state-of-the-art results on MNIST and CIFAR-10 datasets.
• The method demonstrates superior classification accuracy on small-sample sonar image target classification (SITC) datasets.
• Analysis confirms significant advantages in computational complexity and energy consumption for SNNs over ANNs in SITC tasks.

Conclusions:

• The combination of SABP and the SNN framework enables effective direct training of deep SNNs.
• This approach bridges the performance gap between SNNs and ANNs, particularly for specialized tasks.
• Brain-inspired SNNs show strong potential for efficient and high-performance computing in applications like sonar target classification.