Brain-inspired neural circuit evolution for spiking neural networks.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Current spiking neural networks (SNNs) often use deep learning structures, primarily feedforward, limiting their potential on complex tasks.
• Biological neural systems exhibit self-organization and diverse neuron types, forming complex circuits for cognitive functions, a richness not fully captured in current SNN designs.
• Integrating biological neural circuit dynamics into SNN structures remains a significant challenge.

Purpose of the Study:

• To develop a more biologically plausible evolutionary framework for designing SNNs.
• To enhance SNN capabilities by incorporating diverse neural circuit types and biologically inspired learning rules.
• To improve performance in complex tasks like image classification and reinforcement learning.

Main Methods:

• Proposed a novel evolutionary space combining feedforward and feedback connections with excitatory and inhibitory neurons.
• Utilized local spike-timing-dependent plasticity (STDP) and global error signals to adaptively evolve neural circuits (e.g., forward/feedback inhibition, lateral inhibition).
• Implemented the Neural circuit Evolution strategy (NeuEvo) to construct SNNs for image classification and reinforcement learning.

Main Results:

• The evolved SNNs demonstrated significantly enhanced capabilities in perception and reinforcement learning tasks.
• NeuEvo achieved state-of-the-art results on benchmark datasets including CIFAR10, DVS-CIFAR10, DVS-Gesture, and N-Caltech101, and advanced performance on ImageNet.
• When combined with deep reinforcement learning algorithms, the evolved SNNs achieved performance comparable to traditional artificial neural networks.

Conclusions:

• The brain-inspired NeuEvo strategy effectively evolves complex SNNs with rich neural circuit types, overcoming limitations of current SNN design paradigms.
• This approach provides a foundation for creating more sophisticated artificial neural networks inspired by biological systems.
• The evolved spiking neural circuits pave the way for future advancements in complex network evolution for diverse functional applications.