Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Unveiling the Spatial-temporal Effective Receptive Fields of Spiking Neural Networks
Authors: Jieyuan (Eric) Zhang, Xiaolong Zhou, Shuai Wang, Wenjie Wei, Hanwen Liu, Qian Sun, Malu Zhang, Yang Yang, Haizhou Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments conducted on the Meta-SDT variants and across object detection and semantic segmentation tasks further validate the effectiveness of our proposed method. Beyond these specific applications, we believe the proposed ST-ERF framework can provide valuable insights for designing and optimizing SNN architectures across a broader range of tasks. |
| Researcher Affiliation | Academia | 1University of Electronic Science and Technology of China, 2The Chinese University of Hong Kong, Shenzhen , 3Shenzhen Loop Area Institute |
| Pseudocode | No | The paper describes methods using mathematical equations and textual descriptions (e.g., in Section 5.1 and 5.2) but does not include any clearly labeled pseudocode or algorithm blocks. For example, equations (8) and (9) define MLPixer and SRB, and (10) and (11) define SSC, but these are not presented as structured algorithm steps. |
| Open Source Code | Yes | The code is available at Eric Zhang1412/Spatial-temporal-ERF. |
| Open Datasets | Yes | Extensive experiments conducted on the Meta-SDT variants and across object detection and semantic segmentation tasks further validate the effectiveness of our proposed method... on COCO 2017 object detection and ADE20K semantic segmentation... We initialized the central spatial feature across all channels and timesteps in the output tensor as uniform gradient stimuli (value = 1), and propagated the gradients backward through the network. Each experiment consisted of 60 iterations with input tensors randomly drawn from a standard normal distribution (ยต = 0, ฯ2 = 1). The comparison results of object detection and instance segmentation are shown in Table 1. Under the same training schedule, both the MLPixer and SRB variants outperforms the baseline across all metrics. More specifically, the SRB variant exceeds the performance of SDTv3-T and SDTv3-B by 10.42% and 4.26% on the APb 50 metric, while maintaining almost the same model size. In conclusion, our approach demonstrates efficacy in object detection and instance segmentation, setting a new benchmark for COCO dataset in the SNN domain. Performance on ADE20K We evaluate the performance of MLPixer and SRB on the semantic segmentation task using the challenging ADE20K dataset [66]. |
| Dataset Splits | Yes | For COCO 2017 dataset, We utilize the MMDetection [72] framework to implement the existing models and our method. The object detection and instance segmentation framework strictly follows Mask R-CNN, with a training schedule of 1 (12 epochs). We use a total batch size of 4/GPU, utilize the Adam W optimizer with a learning rate of 1 10 4 and a weight decay of 0.05. Images are resized and cropped into 1333 800 for training and testing and maintain the ratio. Random horizontal flipping and resize with a ratio of 0.5 was applied for augmentation during training. This pre-training fine-tuning method is a commonly used strategy in ANNs. For ADE20K dataset, we utilize the MMSegmentation [73] framework. The training configuration strictly encompasses for 160,000 iterations. The batch size is set to 4/GPU, and the Adam W optimizer is used. The learning rate and weight decay parameters are tuned to 2 10 4 and 0.05, respectively. To speed up training, we warm up the model for 1.5k iterations with a linear decay schedule. All the experiments are conducted on 4 NVIDIA-A100 80GB GPUs. |
| Hardware Specification | Yes | All the experiments are conducted on 4 NVIDIA-A100 80GB GPUs. |
| Software Dependencies | No | The paper mentions using "Py Torch's Automatic Differentiation functionality", "MMDetection [72] framework", and "MMSegmentation [73] framework". While these are specific software components, no version numbers are provided for PyTorch, MMDetection, or MMSegmentation, which is necessary for a reproducible description. |
| Experiment Setup | Yes | For COCO 2017 dataset, We utilize the MMDetection [72] framework to implement the existing models and our method. The object detection and instance segmentation framework strictly follows Mask R-CNN, with a training schedule of 1 (12 epochs). We use a total batch size of 4/GPU, utilize the Adam W optimizer with a learning rate of 1 10 4 and a weight decay of 0.05. Images are resized and cropped into 1333 800 for training and testing and maintain the ratio. Random horizontal flipping and resize with a ratio of 0.5 was applied for augmentation during training. This pre-training fine-tuning method is a commonly used strategy in ANNs. For ADE20K dataset, we utilize the MMSegmentation [73] framework. The training configuration strictly encompasses for 160,000 iterations. The batch size is set to 4/GPU, and the Adam W optimizer is used. The learning rate and weight decay parameters are tuned to 2 10 4 and 0.05, respectively. To speed up training, we warm up the model for 1.5k iterations with a linear decay schedule. All the experiments are conducted on 4 NVIDIA-A100 80GB GPUs. Table 5: Hyper-parameters for pre-training on Image Net-1K Hyper-parameter Settings Model size T/M/B Timestemp 4 Epochs 200 Resolution 224*224 Batch size 1568 Optimizer LAMB Base learning rate 6e-4 Learning rate decay Cosine Warmup eopchs 10 Weight decay 0.05 Random augment 9/0.5 Mixup None Cutmix None Label smoothing 0.1 |