Spiking Deep Residual Networks.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking neural networks (SNNs) offer biological plausibility and potential for energy-efficient AI.
• Training very deep SNNs remains a significant challenge.
• Residual Networks (ResNets) are fundamental, state-of-the-art models in deep learning.

Purpose of the Study:

• To propose an efficient method for constructing deep SNNs.
• To adapt successful deep learning architectures like ResNet for SNNs.
• To achieve high performance and low latency in deep SNNs.

Main Methods:

• Conversion of trained artificial neural networks (ANNs), specifically ResNets, into spiking ResNets (S-ResNets).
• Development of a residual conversion model to scale ANN activations to SNN firing rates.
• Implementation of a compensation mechanism to mitigate discretization errors.

Main Results:

• The proposed method successfully created an asynchronous SNN deeper than 100 layers.
• Achieved state-of-the-art performance on CIFAR-10, CIFAR-100, and ImageNet 2012 datasets.
• Demonstrated low latency in the resulting deep SNNs, comparable to original ANNs.

Conclusions:

• The conversion approach enables the creation of high-performance, deep SNNs.
• This method bridges the gap between traditional ANNs and biologically plausible SNNs.
• The work paves the way for more efficient and powerful SNNs in machine intelligence.