Interpretable and Testable Vision Features via Sparse Autoencoders

TL;DR

This paper introduces a method using sparse autoencoders to interpret and test vision model features, enabling semantic understanding and causal probing without retraining models.

Contribution

The authors propose a model-agnostic approach with sparse autoencoders that provides interpretable features and causal control over vision models' learned representations.

Findings

SAEs reveal meaningful semantic differences in pre-trained models

SAEs enable patch-level causal edits across tasks

Method supports interpretability and causal testing without retraining

Abstract

To truly understand vision models, we must not only interpret their learned features but also validate these interpretations through controlled experiments. While earlier work offers either rich semantics or direct control, few post-hoc tools supply both in a single, model-agnostic procedure. We use sparse autoencoders (SAEs) to bridge this gap; each sparse feature comes with real-image exemplars that reveal its meaning and a decoding vector that can be manipulated to probe its influence on downstream task behavior. By applying our method to widely-used pre-trained vision models, we reveal meaningful differences in the semantic abstractions learned by different pre-training objectives. We then show that a single SAE trained on frozen ViT activations supports patch-level causal edits across tasks (classification and segmentation) all without retraining the ViT or task heads. These qualitative, falsifiable demonstrations position SAEs as a practical bridge between concept discovery and causal probing of vision models. We provide code, demos and models on our project website: https://osu-nlp-group.github.io/saev.