One-Hot Multi-Level Leaky Integrate-and-Fire Spiking Neural Networks for Enhanced Accuracy-Latency Tradeoff.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Energy-Efficient Computing

Background:

• Spiking neural networks (SNNs) are promising energy-efficient alternatives to artificial neural networks (ANNs).
• Reducing SNN latency to a single timestep enhances energy efficiency but often degrades accuracy.
• Balancing accuracy and energy consumption in SNNs remains a significant challenge.

Purpose of the Study:

• To introduce a novel neuron model that enhances the accuracy-energy tradeoff in SNNs.
• To explore a new dimension for optimizing SNN performance using a one-hot multi-level leaky integrate-and-fire (M-LIF) neuron.
• To demonstrate the effectiveness of the proposed model for both static and dynamic vision datasets.

Main Methods:

• Developed a novel one-hot multi-level leaky integrate-and-fire (M-LIF) neuron model.
• Represented hidden layer inputs/outputs using one-hot binary-weighted spike lanes.
• Evaluated the model on static image classification (ImageNet) and dynamic vision datasets.

Main Results:

• One-hot M-LIF SNNs achieved 2% higher accuracy than conventional LIF SNNs on ImageNet, with 20x lower energy consumption than ANNs.
• For dynamic vision tasks, M-LIF SNNs reduced latency by 3x compared to conventional LIF SNNs.
• Accuracy degradation was limited to less than 1% for dynamic vision tasks.

Conclusions:

• The one-hot M-LIF neuron model effectively improves the accuracy-energy tradeoff in SNNs.
• This novel approach enables SNNs to achieve superior performance and energy efficiency.
• The M-LIF model offers a viable solution for latency reduction without significant accuracy loss.