Probabilistic Spike Propagation for Efficient Hardware Implementation of Spiking Neural Networks.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Hardware Acceleration

Background:

• Spiking Neural Networks (SNNs) are noted for temporal data processing and low-power hardware potential.
• Optimizing SNN computational efficiency is crucial for resource-limited applications.
• Network complexity correlates with spiking activity, specifically spike propagation.

Purpose of the Study:

• To enhance the computational efficiency of rate-coded Spiking Neural Networks.
• To introduce a novel method for optimizing spike propagation in SNNs.
• To develop specialized hardware for efficient SNN processing.

Main Methods:

• Interpreting synaptic weights as probabilities to regulate spike propagation.
• Developing the Probabilistic Spiking Neural Network Application Processor (P-SNNAP).
• Evaluating the approach on benchmark SNNs.

Main Results:

• Probabilistic spike propagation reduced propagated spikes by 2.4-3.69×.
• The P-SNNAP accelerator achieved 1.39-2× energy reduction.
• Simultaneous speedups of 1.16-1.62× were observed compared to traditional SNN evaluation.

Conclusions:

• Probabilistic spike propagation significantly reduces computational overhead in SNNs.
• The P-SNNAP demonstrates the practical benefits of this approach for energy and speed.
• This method is effective for deploying SNNs in energy-efficient, resource-constrained environments.