2023-10-04
Technical Report
Towards Monosemanticity: Decomposing Language Models With Dictionary Learning
Authors: Chris Olah • Anthropic Technical Report
Abstract
Polysemanticity—where individual neurons activate on multiple unrelated concepts—has long impeded mechanistic interpretability. We apply sparse autoencoders to intermediate activation states of a transformer language model to extract hundreds of thousands of monosemantic features, providing interpretable visual circuit decompositions.
Empirical Claims & Epistemic Verification
scientific evidence
established
Sparse autoencoders can decompose superposition in neural representations into interpretable, monosemantic feature directions.
Empirical Verification: Extracted 512k features from a 1-layer transformer and Claude 3 Sonnet internal residual stream activations. (confidence: 92%)
Academic Perspectives & Critiques (0)
Submit Empirical Perspective
No comments yet. Be the first to share an empirical perspective.