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..

Adaptive Fission: Post-training Encoding for Low-latency Spike Neural Networks

Authors: Yizhou Jiang, Feng Chen, Yihan Li, Yuqian Liu, Haichuan Gao, Tianren Zhang, Ying Fang

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on neuromorphic hardware demonstrate up to 80% reductions in latency and power consumption without degrading accuracy. ... 6 Experiments ... 6.2 Main Results on Image Classification ... 6.3 Main Results on Generation
Researcher Affiliation Collaboration 1Department of Automation, Tsinghua University, Beijing, China 2Beijing Qian Jue Technology Co., Ltd. 3College of Computer and Cyber Security, Fujian Normal University, Fuzhou, China
Pseudocode Yes Algorithm 1 Overall Algorithm Input: Original SNN; time-step T
Open Source Code Yes Code is available at: https://github.com/JiangYizhou16/Adaptive-Fission
Open Datasets Yes For classification, we conduct experiments on CIFAR-100 [26] and Image Net [6] using Res Net20 [18], VGG16 [38], and a spike-based Transformer [49]. ... For image generation on CIFAR-10
Dataset Splits Yes Fission Encoding is applied post-training with 2,000 calibration samples for CIFAR-100 and 20,000 for Image Net. ... calibrated on the full 60,000-sample dataset. ... The baseline models and network architectures used in fission are directly sourced from the original codebase, including the Calibration [28] and QCFS [4] conversion methods, and spike-driven transformers [49].
Hardware Specification Yes Deployments on a Lynxi HP201 neuromorphic chip [34] demonstrate up to 80% reductions in latency and energy consumption without compromising accuracy. ... All models are deployed on a Lynxi HP201 neuromorphic accelerator (16GB), a commercial successor to Tianjic [34] ... All our experiments were conducted on a workstation with Intel-13900KS, 32GB memory with a single 4090 GPU
Software Dependencies No The provided Lyn BIDL library enables the compilation of SNN models based on Py Torch. However, no specific version numbers for PyTorch or Lyn BIDL are provided in the paper text.
Experiment Setup Yes Fission Encoding is applied post-training with 2,000 calibration samples for CIFAR-100 and 20,000 for Image Net. ... Converted models use T = 32 time-steps, while directly trained models use T = 4. ... with a batch size of 32. ... The model consists of 4 residual blocks with a channel configuration of [2, 4, 4, 2], and uses the Swish activation function instead of Re LU.