{"title":"评估开源稀疏自动编码器在 GPT-2 Small 中析取事实知识的能力","authors":"Maheep Chaudhary, Atticus Geiger","doi":"arxiv-2409.04478","DOIUrl":null,"url":null,"abstract":"A popular new method in mechanistic interpretability is to train\nhigh-dimensional sparse autoencoders (SAEs) on neuron activations and use SAE\nfeatures as the atomic units of analysis. However, the body of evidence on\nwhether SAE feature spaces are useful for causal analysis is underdeveloped. In\nthis work, we use the RAVEL benchmark to evaluate whether SAEs trained on\nhidden representations of GPT-2 small have sets of features that separately\nmediate knowledge of which country a city is in and which continent it is in.\nWe evaluate four open-source SAEs for GPT-2 small against each other, with\nneurons serving as a baseline, and linear features learned via distributed\nalignment search (DAS) serving as a skyline. For each, we learn a binary mask\nto select features that will be patched to change the country of a city without\nchanging the continent, or vice versa. Our results show that SAEs struggle to\nreach the neuron baseline, and none come close to the DAS skyline. We release\ncode here: https://github.com/MaheepChaudhary/SAE-Ravel","PeriodicalId":501347,"journal":{"name":"arXiv - CS - Neural and Evolutionary Computing","volume":"51 1","pages":""},"PeriodicalIF":0.0000,"publicationDate":"2024-09-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Evaluating Open-Source Sparse Autoencoders on Disentangling Factual Knowledge in GPT-2 Small\",\"authors\":\"Maheep Chaudhary, Atticus Geiger\",\"doi\":\"arxiv-2409.04478\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"A popular new method in mechanistic interpretability is to train\\nhigh-dimensional sparse autoencoders (SAEs) on neuron activations and use SAE\\nfeatures as the atomic units of analysis. However, the body of evidence on\\nwhether SAE feature spaces are useful for causal analysis is underdeveloped. In\\nthis work, we use the RAVEL benchmark to evaluate whether SAEs trained on\\nhidden representations of GPT-2 small have sets of features that separately\\nmediate knowledge of which country a city is in and which continent it is in.\\nWe evaluate four open-source SAEs for GPT-2 small against each other, with\\nneurons serving as a baseline, and linear features learned via distributed\\nalignment search (DAS) serving as a skyline. For each, we learn a binary mask\\nto select features that will be patched to change the country of a city without\\nchanging the continent, or vice versa. Our results show that SAEs struggle to\\nreach the neuron baseline, and none come close to the DAS skyline. We release\\ncode here: https://github.com/MaheepChaudhary/SAE-Ravel\",\"PeriodicalId\":501347,\"journal\":{\"name\":\"arXiv - CS - Neural and Evolutionary Computing\",\"volume\":\"51 1\",\"pages\":\"\"},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2024-09-05\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"arXiv - CS - Neural and Evolutionary Computing\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/arxiv-2409.04478\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"arXiv - CS - Neural and Evolutionary Computing","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/arxiv-2409.04478","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0
摘要
在机理可解释性方面,一种流行的新方法是在神经元激活上训练高维稀疏自动编码器(SAE),并使用 SAE 特征作为分析的原子单位。然而,关于 SAE 特征空间是否有助于因果分析的证据尚不充分。在这项工作中,我们使用 RAVEL 基准来评估在 GPT-2 small 的隐藏表征上训练出来的 SAE 是否拥有一组特征集,可以分别传递城市在哪个国家和哪个大洲的知识。我们以神经元作为基线,以通过分布式对齐搜索(DAS)学习到的线性特征作为天际线,对 GPT-2 small 的四个开源 SAE 进行了对比评估。对于每种方法,我们都会学习一个二进制掩码,以选择将被修补的特征,从而在不改变大陆的情况下改变一个城市的国家,反之亦然。我们的结果表明,SAE 难以达到神经元基线,而且没有一个能接近 DAS 的天际线。我们在此发布代码:https://github.com/MaheepChaudhary/SAE-Ravel
Evaluating Open-Source Sparse Autoencoders on Disentangling Factual Knowledge in GPT-2 Small
A popular new method in mechanistic interpretability is to train
high-dimensional sparse autoencoders (SAEs) on neuron activations and use SAE
features as the atomic units of analysis. However, the body of evidence on
whether SAE feature spaces are useful for causal analysis is underdeveloped. In
this work, we use the RAVEL benchmark to evaluate whether SAEs trained on
hidden representations of GPT-2 small have sets of features that separately
mediate knowledge of which country a city is in and which continent it is in.
We evaluate four open-source SAEs for GPT-2 small against each other, with
neurons serving as a baseline, and linear features learned via distributed
alignment search (DAS) serving as a skyline. For each, we learn a binary mask
to select features that will be patched to change the country of a city without
changing the continent, or vice versa. Our results show that SAEs struggle to
reach the neuron baseline, and none come close to the DAS skyline. We release
code here: https://github.com/MaheepChaudhary/SAE-Ravel