Constructing Accurate and Efficient Deep Spiking Neural Networks With Double-Threshold and Augmented Schemes

Area of Science:

• Artificial Intelligence
• Neuromorphic Computing
• Machine Learning

Background:

• Artificial neural networks (ANNs) face high power consumption challenges.
• Spiking neural networks (SNNs) offer high efficiency but often lag in recognition accuracy.
• Converting trained ANNs to SNNs is a strategy to bridge this performance gap.

Purpose of the Study:

• To introduce novel conversion methods, TerMapping and AugMapping, for enhancing SNN performance.
• To evaluate the extent to which converted SNNs retain ANN accuracy and SNN efficiency.
• To provide new approaches for integrating ANN techniques into SNNs for improved neuromorphic computing.

Main Methods:

• Proposed TerMapping: A double-threshold scheme extending typical threshold-balancing.
• Proposed AugMapping: Incorporates an augmented spike scheme with a spike coefficient.
• Performance evaluation using MNIST, Fashion-MNIST, and CIFAR10 datasets.

Main Results:

• The double-threshold scheme in TerMapping effectively improves SNN accuracies.
• AugMapping demonstrates significant advantages in constructing accurate, fast, and efficient deep SNNs.
• The proposed methods outperform other state-of-the-art approaches in SNN conversion.

Conclusions:

• TerMapping and AugMapping offer viable solutions for enhancing SNN performance.
• AugMapping is particularly effective for creating high-performing deep SNNs.
• These conversion techniques facilitate the integration of ANNs with SNNs, benefiting applied neuromorphic computing.