Research map / Architectures and efficiency / Attention and transformers
Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers
Tang Li, Yanlin Chen, Mengmeng Ma, Xi Peng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train SAE on neuroscience-motivated probing suite instead of ImageNet (ablation) | SAE for ViT steering · WaterBirds worst group | worst-group accuracy (%): 63.9 -> 98.5 | Improved |
| train SAE on probing suite instead of MSCOCO (ablation) | SAE for ViT steering · WaterBirds worst group | worst-group accuracy (%): 95.4 -> 98.5 | Improved |
| remove background concepts by zeroing SAE activations | CLIP-ViT-B-32 linear probe · WaterBirds worst group | worst-group accuracy (%): 50.3 -> 98.5 | Improved |
| enhance background concepts | CLIP-ViT-B-32 linear probe · WaterBirds worst group | worst-group accuracy (%): 50.3 -> 5.3 | Hurt |
| use probing-suite concept set instead of existing concept sets to interpret SAEs (ablation) | SAE interpretation · Interpretation accuracy Top-30 | accuracy (%): 16.7 -> 54.1 (vs Google-20K) | Improved |
| train SAE on probing suite instead of ImageNet or MSCOCO (ablation) | SAE training data · Interpretation accuracy Top-30 | accuracy (%): 50.7 / 51.2 -> 54.1 (Ours-16K concepts) | Improved |
| curate and deduplicate probing set across four abstraction levels | SAE training images · Concept coverage | coverage efficiency (%/1K images): 0.06 -> 1.26 | Improved |
| curate probing set vs full ImageNet | SAE training images · Concept coverage average | average concept coverage (%): 78.9 -> 80.5 | Mixed |
| use CLIP-based concept reading instead of MLLM summarization | SAE feature auto-interpretation · ImageNet-val IoU | IoU: 0.438 -> 0.432; run time 1h 40min -> 5 min | No clear effect |
| use BatchTopK SAE among five SAE architectures (ablation) | ViT SAE · SAE benchmark on CLIP-ViT-B-32 | reconstruction, monosemanticity, orthogonality, dead neurons | Improved |
| use concept-based heatmaps instead of attribution method | concept localization · VOC2007 Quantus | Point Game / Attribution Localization: 41.9 -> 45.0; 32.3 -> 36.0 | Improved |
Closest papers by meaning
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · similarity 0.90)
- Interpreting vision transformers via residual replacement model (NeurIPS 2025 · similarity 0.89)
- Sparse autoencoders reveal selective remapping of visual concepts during adaptation (ICLR 2025 · similarity 0.88)
- VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set (NeurIPS 2025 · similarity 0.87)
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.87)
- NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
