HSNNM
收藏资源简介:
HSNNM(Hebbian稀疏神经网络模型)数据集与一篇研究论文相关联,该论文提出了一种用于Transformer层的高效、稀疏前馈替代方案。该模型旨在通过稀疏前馈网络设计探索效率与准确性之间的权衡,其核心设计包括top-1叶路由、k-Winners-Take-All(k-WTA)叶内稀疏性、局部Hebbian强度调制规则以及net2net风格的结构增长机制。实验使用TinyStories数据集的10%切片进行训练,模型参数为733万个。结果显示,与密集基线相比,每处理一个令牌的浮点运算次数减少了49.6%(从1425万次降至718万次),验证准确率下降了3.4个百分点(从79.45%降至76.06%)。此外,消融压力测试表明Hebbian机制和k-WTA稀疏性与相对较低的灾难性遗忘相关。数据集主要面向文本生成任务,内容语言为英语,涉及研究主题包括Transformer架构、Hebbian学习、稀疏神经网络、高效人工智能、net2net方法和k-WTA。数据集规模较小(样本数少于1K),采用CC BY 4.0许可证。
HSNNM (Hebbian Sparse Neural Network Model) dataset is associated with a research paper that proposes an efficient, sparse feed-forward alternative for Transformer layers. This model aims to explore the trade-off between efficiency and accuracy through sparse feed-forward network design, with its core designs including top-1 leaf routing, k-Winners-Take-All (k-WTA) intra-leaf sparsity, local Hebbian strength modulation rules, and a net2net-style structural growth mechanism. The experiments were trained using a 10% slice of the TinyStories dataset, with the model having 7.33 million parameters. The results demonstrate that compared to the dense baseline, the floating-point operations per processed token decreased by 49.6% (from 14.25 million to 7.18 million), while the top-1 validation accuracy dropped by 3.4 percentage points (from 79.45% to 76.06%). Additionally, ablation stress tests reveal that the Hebbian mechanism and k-WTA sparsity are correlated with relatively low catastrophic forgetting. This dataset is primarily targeted at text generation tasks, with content in English, covering research topics including Transformer architectures, Hebbian learning, sparse neural networks, efficient artificial intelligence, net2net methods, and k-WTA. The dataset has a small scale with fewer than 1,000 samples, and is licensed under CC BY 4.0.
数据集概述
数据集名称: HSNNM: A Sparse, Dynamically Routed Feed-Forward Alternative for Transformer Layers
许可证: CC-BY-4.0
语言: 英文
任务类别: 文本生成
数据集大小: n<1K(少于1000个样本)
标签: paper, transformer, hebbian-learning, sparse-neural-networks, efficient-ai, net2net, kwta
作者: Kastiel Tjuandra
DOI: 10.5281/zenodo.21271668
GitHub: 可访问Zenodo页面获取链接
摘要
该文档描述了一种名为Hebbian稀疏神经网络模型(HSNNM) 的Transformer前馈层替代方案。该方案用一组稀疏、竞争性路由的子网络(称为"lobes")替换标准稠密子层。设计结合了以下机制:
- top-1 lobe路由
- k-Winners-Take-All(k-WTA)intra-lobe稀疏性
- 局部Hebbian强度调制规则
- net2net风格的结构增长机制
在基于TinyStories数据集10%切片训练的7.33M参数模型上,HSNNM相较于匹配的稠密基线模型(14.25M FLOPs/token vs 7.18M FLOPs/token),每个token的FLOPs减少了49.6%,但验证准确率下降了3.4个百分点(76.06% vs 79.45%)。
在5,000步压力测试的消融实验中,Hebbian机制和k-WTA稀疏性与不含这些机制的变体相比,表现出较低的灾难性遗忘,但该比较使用了与主要结果不同的评估协议,应视为提示性而非结论性。
这些结果作为小规模计算跳过型稀疏FFN设计在效率-准确率权衡上的一个数据点被报告,同时附带其局限性。
引用
bibtex @misc{tjuandra2026hsnnm, title={HSNNM: A Hebbian Sparse Neural Network Model for Efficient Transformer Learning}, author={Tjuandra, Kastiel}, year={2026}, publisher={Zenodo}, doi={10.5281/zenodo.21271668} }





