Going Deeper in Spiking Neural Networks: VGG and Residual Architectures

Going Deeper in Spiking Neural Networks: VGG and Residual Architectures

19 Feb 2019 | Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu and Kaushik Roy
This paper addresses the challenge of applying deep Spiking Neural Networks (SNNs) to complex visual recognition tasks, such as CIFAR-10 and ImageNet. The authors propose a novel algorithmic technique, SPIKE-NORM, to generate deep SNNs with VGG and Residual network architectures, achieving significantly better accuracy compared to state-of-the-art methods. The technique involves weight normalization and threshold balancing during the conversion from Artificial Neural Networks (ANNs) to SNNs, ensuring minimal loss in classification accuracy. The paper also explores the impact of sparsity in SNNs, demonstrating reduced hardware overhead due to event-driven computations. The results show that the proposed method achieves competitive performance on both CIFAR-10 and ImageNet datasets, with the best reported classification error rate on the CIFAR-10 dataset being 8.45%. For ImageNet, the top-1 error rate is 30.04% for VGG-16 and 34.53% for ResNet-34. The paper further discusses architectural constraints and design choices for achieving near-lossless conversion, highlighting the importance of considering the actual SNN operation during the conversion process.This paper addresses the challenge of applying deep Spiking Neural Networks (SNNs) to complex visual recognition tasks, such as CIFAR-10 and ImageNet. The authors propose a novel algorithmic technique, SPIKE-NORM, to generate deep SNNs with VGG and Residual network architectures, achieving significantly better accuracy compared to state-of-the-art methods. The technique involves weight normalization and threshold balancing during the conversion from Artificial Neural Networks (ANNs) to SNNs, ensuring minimal loss in classification accuracy. The paper also explores the impact of sparsity in SNNs, demonstrating reduced hardware overhead due to event-driven computations. The results show that the proposed method achieves competitive performance on both CIFAR-10 and ImageNet datasets, with the best reported classification error rate on the CIFAR-10 dataset being 8.45%. For ImageNet, the top-1 error rate is 30.04% for VGG-16 and 34.53% for ResNet-34. The paper further discusses architectural constraints and design choices for achieving near-lossless conversion, highlighting the importance of considering the actual SNN operation during the conversion process.
Reach us at info@study.space