跳到主要内容Skip to main content

稀疏自编码器Sparse autoencoder

Sparse autoencoder

  • 核心概念Core concepts
  • 第四站 · 看懂Station 4 · Under the Hood
  • 引用核验 2026-11-19refs checked 2026-11-19
解决什么问题What problem it solves

把模型内部密密麻麻的激活拆成一组稀疏、可命名的特征方向。Split dense model activations into a sparse set of nameable feature directions.

这是什么What it is

在模型某层的激活上训练一个宽得多的自编码器。

加稀疏惩罚,让每个输入只激活少量隐藏单元。

这些隐藏单元就是候选的「单义特征」,比原始神经元更容易解释。

Train a much wider autoencoder on activations from one layer.

Add a sparsity penalty so each input activates only a few hidden units.

Those units become candidate monosemantic features that are easier to interpret.

为什么需要知道Why it matters

原始神经元经常一个神经元管多个概念(多义性)。稀疏自编码器是可解释性研究的主流工具。Raw neurons are often polysemantic. Sparse autoencoders are a mainstream interpretability tool.

出处Sources
  • Towards Monosemanticity(Anthropic)Towards Monosemanticity: Decomposing Language Models With Dictionary Learning (Anthropic) ↗