Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers

2026-06-04

AI Summary

Vision Transformers (ViTs) often rely on spurious cues, undermining their interpretability and controllability and thereby hindering safe deployment. To address this, this work proposes ViSAE, a neuroscientifically inspired toolbox that constructs a probing framework integrating 64K images with a vocabulary of 16K visual concepts. It introduces top-down concept reading and bottom-up circuit tracing algorithms to enable efficient analysis and intervention in ViT internal mechanisms. Leveraging sparse autoencoders, a large-scale concept lexicon, and concept editing techniques, the method achieves a 20-fold improvement in concept coverage efficiency over ImageNet, a 28.7% gain in explanation accuracy, and a 48.2% increase in worst-group accuracy on the WaterBirds dataset, substantially outperforming existing approaches.

Abstract

Despite high accuracy, Vision Transformer (ViT) predictions can be driven by spurious cues, raising the need to understand their inner workings before safe deployment. Sparse autoencoders (SAEs) provide a promising lens for decomposing model representations into human-interpretable concepts, yet adapting SAE-based interpretation to ViTs remains challenging due to limited control over concept coverage and subjective, non-scalable feature interpretation. To fill the gaps, motivated by neuroscience-inspired principles, we propose ViSAE, a mechanistic interpretability toolbox for understanding ViT inner workings through concept circuits. ViSAE consists of three components: (1) A probing suite with 64K images and a 16K visually grounded concept vocabulary, improving concept coverage efficiency by 20x over ImageNet and interpretation accuracy by 28.7% over existing concept sets. (2) Top-down concept reading and Bottom-up circuit tracing algorithms that automatically recover ViT inner workings via concept circuits. (3) Applications for auditing and steering ViT behavior. Through concept editing, ViSAE improves the worst-group accuracy on WaterBirds by 48.2%, outperforming existing methods by 23.8%. Our data and code: https://github.com/deep-real/ViSAE.

Problem

• Vision Transformer
• interpretability
• spurious cues
• concept coverage
• feature interpretation

Innovation

• Vision Transformers
• mechanistic interpretability
• concept circuits
• sparse autoencoders
• neuroscience-inspired