Brain-inspired chaotic spiking backpropagation.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) offer high energy efficiency by mimicking biological neural systems but are challenging to train effectively.
• Current training methods, like surrogate gradients, often lead to SNNs getting stuck in local minima due to reliance on gradient dynamics.

Purpose of the Study:

• To introduce a novel training method, chaotic spiking backpropagation (CSBP), for improved SNN performance.
• To leverage brain-inspired chaotic dynamics to enhance the effectiveness and robustness of SNN learning.

Main Methods:

• Developed CSBP, incorporating a loss function that generates brain-like chaotic dynamics during training.
• Utilized the ergodic and pseudo-random properties of chaotic dynamics to facilitate SNN learning.
• Analyzed the theoretical learning process, observing initial chaos followed by bifurcations and convergence to gradient dynamics.

Main Results:

• CSBP significantly outperformed state-of-the-art methods on neuromorphic (DVS-CIFAR10, DVS-Gesture) and large-scale static datasets (CIFAR100, ImageNet).
• Demonstrated superior accuracy and robustness in trained SNNs using the CSBP method.
• Theoretical analysis confirmed a learning trajectory mirroring biological brain activity patterns.

Conclusions:

• CSBP provides a powerful new tool for direct and effective training of spiking neural networks.
• The method enhances SNN performance by utilizing chaotic dynamics, addressing limitations of traditional gradient-based approaches.
• Offers valuable insights into biological brain learning mechanisms through computational modeling.