Noise Exploitation for ANN-to-SNN Conversion
Injecting controllable noise during ANN quantization training to model SNN spike dynamics, enabling seamless end-to-end ANN-to-SNN conversion.
The Noise Adaptor injects controllable noise during ANN quantization training to model SNN spike dynamics. This enables seamless end-to-end conversion without runtime noise correction.
Key Results:
- CIFAR-10 (ResNet-18): 95.26% (T=2), 95.95% (T=4), 96.61% (T=8), 96.72% (T=16)
- ImageNet (ResNet-50): 37.08% (T=2), 58.76% (T=4), 71.50% (T=8), 75.69% (T=16)
Publication: Li C, Rajendran B. Noise Adaptor in Spiking Neural Networks. arXiv 2024.