← 研踪 PaperTrailarXiv:2605.31304 · v2
DUNE: Disentangling Polysemantic Neurons Without Losing Predictive Performance
首次提交 2026-05-29 21:39
版本更新 2026-11-12 06:57
雷达观测 2026-11-13 14:16
摘要
Deep neural networks (DNNs) are widely used, but understanding what they actually learn remains difficult. A major obstacle is that individual neurons often respond to multiple unrelated concepts, obscuring the network's internal decision process. Existing approaches, such as sparse autoencoders, can disentangle these mixed signals into more meaningful, ``monosemantic'' features, but typically introduce approximation errors that may degrade downstream performance. Consequently, these methods exhibit a Pareto frontier between monosemanticity and functional faithfulness. To address this, we introduce DUNE (disentangling units with no error), a method that decomposes DNN neurons into more monosemantic subunits while preserving their original function by design. DUNE identifies concept-specific contributions from the preceding layer and re-routes them into separate subunits whose aggregation exactly recovers the original activation. DUNE can be applied directly to pretrained models, without labels or training of an auxiliary model. Across several models, including DINOv2, DUNE achieves state-of-the-art or competitive monosemanticity while exactly preserving functional faithfulness, substantially advancing the existing Pareto frontier between these properties. DUNE operates efficiently and enables fine-grained interventions such as concept-specific steering.
作者备注：Preprint
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-12 06:57 · 最新观测版本
DUNE: Disentangling Polysemantic Neurons Without Losing Predictive Performance
Deep neural networks (DNNs) are widely used, but understanding what they actually learn remains difficult. A major obstacle is that individual neurons often respond to multiple unrelated concepts, obscuring the network's internal decision process. Existing approaches, such as sparse autoencoders, can disentangle these mixed signals into more meaningful, ``monosemantic'' features, but typically introduce approximation errors that may degrade downstream performance. Consequently, these methods exhibit a Pareto frontier between monosemanticity and functional faithfulness. To address this, we introduce DUNE (disentangling units with no error), a method that decomposes DNN neurons into more monosemantic subunits while preserving their original function by design. DUNE identifies concept-specific contributions from the preceding layer and re-routes them into separate subunits whose aggregation exactly recovers the original activation. DUNE can be applied directly to pretrained models, without labels or training of an auxiliary model. Across several models, including DINOv2, DUNE achieves state-of-the-art or competitive monosemanticity while exactly preserving functional faithfulness, substantially advancing the existing Pareto frontier between these properties. DUNE operates efficiently and enables fine-grained interventions such as concept-specific steering.
查看此版本解读 →查看此版本原文 ↗