Efficient ANN-SNN Conversion with Error Compensation Learning

Authors: Chang Liu, Jiangrong Shen, Xuming Ran, Mingkun Xu, Qi Xu, Yi Xu, Gang Pan

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on CIFAR-10, CIFAR-100, Image Net datasets show that our method achieves high-precision and ultra-low latency among existing conversion methods.
Researcher Affiliation Academia 1School of Computer Science and Technology, Dalian University of Technology, Dalian, China 2Faculty of Electronic and Information Engineering, Xi an Jiaotong University 3State Key Lab of Brain-Machine Intelligence, Zhejiang University 4National University of Singapore 5Guangdong Institute of Intelligence Science and Technology, Zhuhai, China 6College of Computer Science and Technology, Zhejiang University. Correspondence to: Qi Xu <EMAIL>.
Pseudocode No The paper describes methods and equations but does not contain explicitly structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statement about releasing source code or a link to a code repository.
Open Datasets Yes Experimental results on CIFAR-10, CIFAR-100, Image Net datasets show that our method achieves high-precision and ultra-low latency among existing conversion methods.
Dataset Splits No The paper mentions using CIFAR-10, CIFAR-100, and Image Net datasets but does not explicitly provide details about training/test/validation splits.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory amounts) used for conducting the experiments.
Software Dependencies No The paper does not provide specific software dependency details, such as library names with version numbers, needed to replicate the experiment.
Experiment Setup Yes In our dual threshold neuron method, the quantization steps L is a hyperparameter that affects the accuracy of the converted SNN. To better understand the impact of L on SNN performance and determine the optimal value, we trained VGG16, Res Net-20, and Res Net-18 networks with a pruning function with a learnable threshold λ using different quantization steps L, including 2, 4, 8, 16, and 32, and then converted them to SNNs. ... this paper sets the negative threshold to a small negative value (-1e-3 according to experience).