The Linear Centroids Hypothesis: Features as Directions Learned by Local Experts

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Walker, Thomas, Humayun, Ahmed Imtiaz, Balestriero, Randall, Baraniuk, Richard
Format: Preprint
Publié: 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866910199141367808
author Walker, Thomas
Humayun, Ahmed Imtiaz
Balestriero, Randall
Baraniuk, Richard
author_facet Walker, Thomas
Humayun, Ahmed Imtiaz
Balestriero, Randall
Baraniuk, Richard
contents The Linear Representation Hypothesis (LRH) identifies features of a trained deep network (DN) as linear directions in the activation spaces, i.e., output spaces of intermediate layers. This characterization decouples the input-output maps learned by a DN from the organization of feature directions in its activation spaces. We introduce the Linear Centroids Hypothesis (LCH), which instead identifies features with linear directions among a DN's centroid spaces -- where any vector denotes a centroid or summary of a local affine expert characterizing the learned input-output maps of the DN exactly (e.g., for piecewise-affine DNs) or approximately (e.g., for smooth DNs like transformers). We show that replacing intermediate activations with centroids yields a functional drop-in alternative for standard interpretability tools. Empirically, this change yields sparser, more downstream-useful feature dictionaries on DINO ViTs, suppresses spurious directions on a controlled task, recovers interpretable circuits in GPT2-Large, and produces faithful gradient-based saliency maps. LCH unifies dictionaries, probing, circuits, and saliency maps into a single geometric object grounded in the network's input-output map -- making interpretability mechanistic by construction rather than post hoc. Code to study the LCH https://github.com/ThomasWalker1/LinearCentroidsHypothesis .
format Preprint
id arxiv_https___arxiv_org_abs_2604_11962
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle The Linear Centroids Hypothesis: Features as Directions Learned by Local Experts
Walker, Thomas
Humayun, Ahmed Imtiaz
Balestriero, Randall
Baraniuk, Richard
Machine Learning
The Linear Representation Hypothesis (LRH) identifies features of a trained deep network (DN) as linear directions in the activation spaces, i.e., output spaces of intermediate layers. This characterization decouples the input-output maps learned by a DN from the organization of feature directions in its activation spaces. We introduce the Linear Centroids Hypothesis (LCH), which instead identifies features with linear directions among a DN's centroid spaces -- where any vector denotes a centroid or summary of a local affine expert characterizing the learned input-output maps of the DN exactly (e.g., for piecewise-affine DNs) or approximately (e.g., for smooth DNs like transformers). We show that replacing intermediate activations with centroids yields a functional drop-in alternative for standard interpretability tools. Empirically, this change yields sparser, more downstream-useful feature dictionaries on DINO ViTs, suppresses spurious directions on a controlled task, recovers interpretable circuits in GPT2-Large, and produces faithful gradient-based saliency maps. LCH unifies dictionaries, probing, circuits, and saliency maps into a single geometric object grounded in the network's input-output map -- making interpretability mechanistic by construction rather than post hoc. Code to study the LCH https://github.com/ThomasWalker1/LinearCentroidsHypothesis .
title The Linear Centroids Hypothesis: Features as Directions Learned by Local Experts
topic Machine Learning
url https://arxiv.org/abs/2604.11962