High-performance deep spiking neural networks via at-most-two-spike exponential coding.

Area of Science:

• Neuromorphic Computing
• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking neural networks (SNNs) are crucial for efficient neuromorphic computing.
• Converting advanced artificial neural networks (ANNs) to SNNs is a key strategy for high-performance SNN development.
• Existing conversion methods face challenges in balancing accuracy, latency, and power consumption.

Purpose of the Study:

• To propose a novel ANN to SNN conversion methodology using a time-based coding scheme.
• To introduce the At-most-two-spike Exponential Coding (AEC) scheme and a corresponding AEC spiking neuron model.
• To enhance the accuracy, energy efficiency, and inference latency of converted SNNs.

Main Methods:

• Developed the At-most-two-spike Exponential Coding (AEC) scheme with a novel spiking neuron model.
• Utilized two exponential decay functions for dynamic encoding thresholds to represent pixel intensities.
• Fine-tuned AEC neuron hyper-parameters using a loss function and introduced regularization terms for spike count.

Main Results:

• The AEC method achieves higher accuracy in deep SNNs compared to existing conversion techniques.
• Demonstrated significant improvements in energy efficiency for SNN inference.
• Showcased reduced inference latency in SNNs converted using the AEC scheme.

Conclusions:

• The proposed AEC conversion methodology offers a superior approach for developing high-performance SNNs.
• AEC effectively balances accuracy, latency, and power consumption, making it suitable for neuromorphic applications.
• This work provides a valuable contribution to the field of efficient SNN implementation.