Effective Plug-Ins for Reducing Inference-Latency of Spiking Convolutional Neural Networks During Inference Phase.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel in classification, while Spiking Neural Networks (SNNs) offer energy efficiency through event-driven processing.
• Existing research has focused on weight normalization, training adjustments, or architectural changes to reduce SNN inference latency.
• Little attention has been given to optimizing the inference phase itself for SCNNs.

Purpose of the Study:

• To propose and evaluate novel, low-cost stopping criteria for reducing the inference latency of Spiking Convolutional Neural Networks (SCNNs).
• To investigate the impact of these criteria on inference speed and classification accuracy.
• To provide a new perspective on SCNN optimization by focusing on the end of the inference phase.

Main Methods:

• Developed four distinct stopping criteria designed as plug-ins for SCNNs.
• Validated the proposed methods using Spiking-AlexNet on CIFAR-10 and Spiking-LeNet-5 on MNIST datasets.
• Implemented simulations on MATLAB and PyTorch platforms to assess performance.

Main Results:

• The proposed methods reduced Spiking-AlexNet inference latency by 3.34 times (from 892 to 267 time steps) with a minor accuracy drop (87.95% to 87.72%).
• Spiking-LeNet-5 models achieved 24-70 time steps per image, a 1.92 to 3.21 times speedup compared to models without the criteria, with accuracy decline under 0.1%.
• Demonstrated significant latency reduction with negligible impact on classification accuracy.

Conclusions:

• The proposed stopping criteria effectively reduce SCNN inference latency.
• These plug-in methods offer a practical and efficient approach to accelerate SCNNs without compromising accuracy.
• The findings open new avenues for optimizing SNNs by focusing on inference-time strategies.