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..
S$^2$NN: Sub-bit Spiking Neural Networks
Authors: Wenjie Wei, Malu Zhang, Jieyuan (Eric) Zhang, Ammar Belatreche, Shuai Wang, Yimeng Shan, Hanwen Liu, Honglin Cao, Guoqing Wang, Yang Yang, Haizhou Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive results on vision tasks reveal that S2NN outperforms existing quantized SNNs in both performance and efficiency, making it promising for edge computing applications. Extensive experiments demonstrate that integrating OS-Quant and MPFD into the baseline enables S2NN to achieve state-of-the-art (SOTA) performance and efficiency. In this section, we evaluate the performance of S2NN on various tasks, including classification, object detection, and semantic segmentation. Then, we conduct ablation studies to verify the effect of the OS-Quant and MPFD. |
| Researcher Affiliation | Academia | Wenjie Wei1, Malu Zhang1,2 , Jieyuan Zhang1, Ammar Belatreche3, Shuai Wang1, Yimeng Shan1, Hanwen Liu1, Honglin Cao1, Guoqing Wang1, Yang Yang1, Haizhou Li2,5 1University of Electronic Science and Technology of China, 2Shenzhen Loop Area Institute, 3Northumbria University, 4Liaoning Technical University 5The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen) Corresponding author: EMAIL |
| Pseudocode | Yes | Algorithm 1 One training iteration process of the S2NN. |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We provide the relevant information in the appendix. |
| Open Datasets | Yes | Dataset CIFAR-10 [72] is a widely used computer vision dataset... CIFAR-100 [72] maintains identical image dimensions... Image Net-1K [73] is a large-scale visual database... DVS-CIFAR10 [74] consists of 10,000 event streams... COCO 2017 [76] is a large-scale computer vision dataset... ADE20K is a comprehensive semantic segmentation dataset... |
| Dataset Splits | Yes | Image Net-1K [73] is a large-scale visual database comprising over 1.2 million training images and 50,000 validation images across 1,000 object categories. COCO 2017 [76] is a large-scale computer vision dataset... The dataset consists of 118,287 training images, 5,000 validation images, and 40,670 test images. |
| Hardware Specification | Yes | We compare S2NN and BSNN on an FPGA to quantify S2NN s advantages. Experimental Settings: Platform: Xilinx Vivado 2021.2 Simulation Platform: Modelsim Clock Frequency: 100MHz AXI Bit Width: 32 bits Model: SCNN: 32x32-64c3-128c3-128c3-256c3-256c3-256c3-10, T=8, η=5 |
| Software Dependencies | Yes | Platform: Xilinx Vivado 2021.2 Simulation Platform: Modelsim |
| Experiment Setup | Yes | Table 12: Hyper-parameters for image classification. Hyper-parameter Image Net-1K CIFAR-10 CIFAR-100 DVS-CIFAR10 Timestep 1 4 6 6 10 Epochs 200 250 250 300 Resolution 224 224 32 32 32 32 48 48 Batch size 1024 128 128 32 Optimizer Adam Adam Adam Adam Weight decay 0 0 0 0 Initial learning rate 6e-4 5e-4 5e-4 5e-4 Learning rate decay Cosine Cosine Cosine Cosine Warmup epochs 10 None None None Label smoothing 0.1 None None None |