Structure Adaptive Control with Neuron Pruning-Growing Strategy for Spiking Neural Networks

Developing efficient adaptive learning techniques for spiking neural networks (SNNs) presents a challenge, primarily due to the unique encoding of spike trains and their implicit nonlinear dynamics. This study introduces a new approach for constructing the dynamic structure of multilayer SNNs aimed at complex spa-tiotemporal pattern learning. By combining spike train learning with a pruning-and-growing strategy, we propose a Structure Adaptive Control (SAC) method to address a key challenge in neural network design: improving network functionality through a flexible, adaptable structure rather than a fixed one. The primary innovation of this research is its capability to dynamically modify the connection weights and neurons within the hidden layers, allowing the network architecture to evolve in response to the similarity of spike trains. The proposed method is eemployed to medical pattern classification tasks, and the impact of three different neuron-adding strategies on algorithm performance is analyzed. The results show that networks with adaptive structures outperform fixed-structure networks in spatiotemporal pattern learning. The findings from the studies substantiate the performance of the SAC method for multilayer SNNs.