Optimal ANN-SNN Conversion for High-accuracy and Ultra-low-latency Spiking Neural Networks
Authors: Tong Bu, Wei Fang, Jianhao Ding, PENGLIN DAI, Zhaofei Yu, Tiejun Huang
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on CIFAR-10/100 and Image Net datasets, and show that it outperforms the stateof-the-art ANN-SNN and directly trained SNNs in both accuracy and time-steps. |
| Researcher Affiliation | Academia | Tong Bu1, Wei Fang1, Jianhao Ding1, Peng Lin Dai2, Zhaofei Yu1 *, Tiejun Huang1 1 Peking University, 2 Southwest Jiaotong University |
| Pseudocode | Yes | Algorithm 1 Algorithm for ANN-SNN conversion. |
| Open Source Code | Yes | Code is available at https://github.com/putshua/SNN conversion QCFS |
| Open Datasets | Yes | CIFAR-10 (Le Cun et al., 1998), CIFAR-100 (Krizhevsky et al., 2009), and Image Net datasets (Deng et al., 2009). |
| Dataset Splits | No | The paper provides training and test set sizes for CIFAR-10, CIFAR-100, and ImageNet, but it does not specify explicit validation dataset splits (e.g., specific percentages or sample counts for a validation set). |
| Hardware Specification | No | The paper mentions neuromorphic hardware in general context and cites power consumption baselines from other works (e.g., ROLLS neuromorphic processor), but it does not specify the particular hardware (e.g., GPU, CPU models, or cloud instances) used to run the authors' experiments. |
| Software Dependencies | No | The paper mentions software platforms like TensorFlow and PyTorch and optimizers like Stochastic Gradient Descent (with a citation), but it does not specify version numbers for any software components, libraries, or dependencies used in their experiments. |
| Experiment Setup | Yes | We use the Stochastic Gradient Descent optimizer (Bottou, 2012) with a momentum parameter of 0.9. The initial learning rate is set to 0.1 for CIFAR-10 and Image Net, and 0.02 for CIFAR-100. A cosine decay scheduler (Loshchilov & Hutter, 2016) is used to adjust the learning rate. We apply a 5e-4 weight decay for CIFAR datasets while applying a 1e-4 weight decay for Image Net. We train all models for 300 epochs. The quantization steps L is set to 4 when training all the networks on CIFAR-10, and VGG-16, Res Net-18 on CIFAR-100 dataset. When training Res Net-20 on CIFAR-100, the parameter L is set to 8. When training Res Net-34 and VGG-16 on Image Net, the parameter L is set to 8, 16, respectively. |