Exploring the Connection Between Binary and Spiking Neural Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Edge intelligence demands reduced computational load for machine learning.
• Binary Neural Networks (BNNs) and Spiking Neural Networks (SNNs) offer solutions but have synergies unexplored.
• Extreme quantization in neural networks is key for efficient on-chip processing.

Purpose of the Study:

• To explore the synergy between Binary Neural Networks and Spiking Neural Networks.
• To investigate the feasibility of training SNNs in an extreme quantization regime.
• To enable Binary Spiking Neural Networks on existing hardware accelerators.

Main Methods:

• Training SNNs using techniques from standard non-spiking networks via a conversion process.
• Evaluating performance on large-scale datasets (CIFAR-100, ImageNet).
• Applying design-time and run-time optimizations to reduce inference latency.

Main Results:

• Achieved near full-precision accuracy with SNNs in the extreme quantization regime.
• Demonstrated that BSNNs can be deployed on BNN-specific "In-Memory" hardware without accuracy loss.
• Reduced inference latency of spiking networks by an order of magnitude compared to prior work.

Conclusions:

• Extreme quantization of SNNs is a viable path to high-accuracy, low-compute edge intelligence.
• The proposed conversion method allows leveraging existing BNN hardware for advanced SNN applications.
• Optimizations significantly enhance the efficiency of both binary and full-precision SNNs for real-time inference.