Low-Latency Spiking Neural Networks Using Pre-Charged Membrane Potential and Delayed Evaluation.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking Neural Networks (SNNs) offer biological plausibility and event-driven computation.
• Converting trained Analog-Valued Neural Networks (ANNs) to SNNs yields high performance.
• SNNs exhibit inherent latency due to temporal integration and spike-based encoding, hindering practical application, especially in deep networks.

Purpose of the Study:

• To propose novel methods for reducing the inherent latency in Spiking Neural Networks.
• To maintain or improve the accuracy of SNNs while reducing latency.
• To enhance the energy efficiency of SNNs by reducing the number of spikes required for stable performance.

Main Methods:

• Introduction of a pre-charged membrane potential (PCMP) mechanism.
• Development of a delayed evaluation (DE) method to discard initial transient errors.
• Application and conversion of neural network models (classification, autoencoder) trained on MNIST and CIFAR-10 datasets to SNNs.

Main Results:

• Significant reduction in SNN latency achieved using PCMP without accuracy loss.
• Further latency reduction demonstrated by the delayed evaluation (DE) method.
• PCMP and DE can be combined for enhanced latency reduction.
• Improved efficiency in the number of spikes required for steady-state performance, beneficial for energy-efficient computing.

Conclusions:

• PCMP and DE are effective strategies for mitigating latency in Spiking Neural Networks.
• The proposed methods enable SNNs to reach steady-state performance faster and with fewer spikes.
• These advancements contribute to the practical applicability and energy efficiency of SNNs in various computational tasks.