Training Deep Convolutional Spiking Neural Networks With Spike Probabilistic Global Pooling.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking neural networks (SNNs) are increasingly explored for deep architectures, often trained with backpropagation (BP).
• Direct BP training of SNNs is computationally intensive and involves numerous trainable parameters, risking overfitting.

Purpose of the Study:

• To present a novel training method, spike probabilistic global pooling (SPGP), for deep convolutional SNNs (DSNNs).
• To simplify the training of DSNNs by reducing the number of trainable parameters and mitigating overfitting.

Main Methods:

• Utilizing a probability function inspired by global pooling in convolutional neural networks (CNNs).
• Employing the discrete leaky-integrate-fire model and spatiotemporal BP algorithm for direct DSNN training.
• Implementing the proposed SPGP method to manage parameters in multi-layered DSNNs.

Main Results:

• SPGP-trained DSNNs achieve competitive performance on image and neuromorphic datasets.
• The method effectively minimizes trainable parameters compared to existing DSNNs.
• Demonstrated improvements in performance, convergence speed, and generalization ability.

Conclusions:

• The SPGP method offers an effective approach for training deep SNNs.
• SPGP successfully addresses the challenges of parameter complexity and overfitting in DSNNs.
• This technique enhances the practical applicability and performance of SNNs in various domains.