Efficient Deep Spiking Multilayer Perceptrons With Multiplication-Free Inference.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Computer Vision

Background:

• Deep convolution architectures for Spiking Neural Networks (SNNs) improve image classification and reduce computation.
• Multiplication-free inference (MFI) in SNNs struggles with attention/transformer mechanisms crucial for high-resolution vision.
• Existing SNNs face limitations in integrating global and local feature extraction effectively.

Purpose of the Study:

• To develop an efficient spiking MLP architecture compatible with multiplication-free inference (MFI).
• To enhance local feature extraction and integrate global receptive fields in SNNs.
• To improve image classification performance on high-resolution vision tasks using SNNs.

Main Methods:

• Proposed an innovative spiking MLP architecture incorporating batch normalization (BN) for MFI compatibility.
• Introduced a spiking patch encoding (SPE) layer for improved local feature extraction.
• Developed an efficient multistage spiking MLP network for comprehensive spike-based computation.

Main Results:

• Achieved 66.39% top-one accuracy on ImageNet-1K, outperforming spiking ResNet-34 by 2.67% without pretraining.
• Reduced computational costs, model parameters, and simulation steps compared to existing SNNs.
• An expanded network variant reached 71.64% top-one accuracy with 2.1x smaller model capacity than spiking VGG-16.

Conclusions:

• The proposed deep SNN architecture effectively integrates global and local learning abilities for superior performance.
• This approach offers a promising pathway for efficient and high-performance spike-based image classification.
• The network's trained receptive fields exhibit activity patterns similar to cortical cells, suggesting biological plausibility.