An all integer-based spiking neural network with dynamic threshold adaptation.

Area of Science:

• Artificial Intelligence
• Computer Science
• Neuroscience

Background:

• Spiking Neural Networks (SNNs) offer energy efficiency but face challenges in achieving high accuracy and fast inference.
• Traditional ANN to SNN conversion methods often suffer from spiking synchronization errors.

Purpose of the Study:

• To propose a novel SNN modeling algorithm for improved accuracy and inference speed.
• To address spiking synchronization errors in SNNs.
• To enhance hardware compatibility and energy efficiency of SNNs.

Main Methods:

• Developed a sophisticated SNN algorithm with a dynamic threshold adaptation mechanism.
• Quantized all SNN variables (membrane potential, threshold, synaptic weights) to integers.
• Implemented spiking LeNet and VGG-Net architectures.

Main Results:

• Achieved >99.45% accuracy on MNIST and >93.15% on CIFAR-10 with only 4 and 8 time steps, respectively.
• Integer quantization significantly reduced computational operations.
• Demonstrated high compatibility with hardware implementation.

Conclusions:

• The proposed SNN algorithm effectively overcomes limitations of traditional methods.
• Integer-based quantization and dynamic threshold adaptation lead to high accuracy and efficiency.
• The approach holds significant potential for energy-efficient edge computing applications.