Research map / Vision-language and multimodal / Human and embodied modeling
Representational Similarity via Interpretable Visual Concepts
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RSVC, which compares two vision networks by extracting CRAFT concepts from each and measuring how well one model's activations regress onto the other's concept coefficients, yielding interpretable per-concept similarity scores.
Problem
Existing representational similarity methods (CCA, CKA, RSA, Brain-score) give a single number for two networks' similarity and do not explain what makes them similar or dissimilar, or which concepts drive differences in model decisions.
Approach
Extend concept-based XAI (CRAFT dictionary learning) to compare two models: extract concepts per model, then use sparse (lasso) regression from one model's activations to the other model's concept coefficients, scoring per-concept cross-model similarity, paired with concept integrated gradients importance and a replacement test linking similarity to behavior.
Key results
- Cross-model concept similarity for most concepts is medium/high, with dense regions above 0.6 similarity across the four model pairs.
- In the toy pink-square experiment, the model trained to ignore the concept has near-zero similarity to the concept that fixates on the pink square.
- DINO vs. MAE fine-tuned on NABirds yields an average bidirectional similarity of 0.80, while DINO seed variants reach 0.89.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with a different seed | ResNet-18 trained on NABirds · NABirds | concept similarity | No clear effect |
| exclude 169 waterbird classes during training, then freeze the backbone and train the classification head on the full dataset | ResNet-18 trained on NABirds · NABirds | classification accuracy: 34% -> 25% | Hurt |
| exclude 169 waterbird classes during training | ResNet-18 trained on NABirds · NABirds | number of dissimilar concepts | Mixed |
| train the backbone on Stanford Cars in addition to NABirds | ResNet-18 trained on NABirds · NABirds and Stanford Cars | concept dissimilarity | Helped |
| finetune on NABirds from ImageNet pretraining | ResNet-18 trained on NABirds · NABirds | concept dissimilarity | Helped |
| vary the finetuning seed | DINO ViT-B and MAE ViT-B on NABirds · NABirds | average concept similarity: 0.80 -> 0.80 | No clear effect |
| compare two DINO models finetuned with different seeds rather than different pretraining strategies | DINO ViT-B on NABirds · NABirds | average concept similarity: 0.80 -> 0.89 | Helped |
| replace coefficients of low-importance concepts | RSVC replacement test · classification task | KL-divergence of logits | No clear effect |
| replace coefficients of high-importance concepts with less-similar predicted coefficients | RSVC replacement test · classification task | KL-divergence of logits | Hurt |
| train a model to use the pink-square concept rather than ignore it | ResNet-18 trained from scratch on modified NABirds · Common Eider concepts | similarity of the pink-square concept | Mixed |
| use regression-based Pearson correlation instead of direct coefficient-matrix correlation (MCS) | RSVC concept comparison · penultimate layers of four model pairs | concept similarity | Helped |
Surprising findings
- Except for DINO vs. MAE, several low/medium similarity concepts still have medium/high importance to model decisions.
- Training a ResNet-18 without waterbird classes leads to only a slight increase in dissimilar concepts despite a large accuracy drop (34% to 25%).
Limitations
- Limited to computer vision models; applicability to LLMs left to future work.
- Inherits the fidelity-interpretability trade-off of concept-based XAI; reconstruction error varies across models.
- Challenging to compare models trained on the same or similar datasets such as ImageNet.
Experimental setup
Datasets: ImageNet NABirds Stanford Cars
Benchmarks: ImageNet NABirds
Metrics: Pearson correlation Spearman correlation concept importance (concept integrated gradients) l2-distance KL-divergence match accuracy
Baselines: CCA CKA RSA Brain-score Rosetta Neurons
Builds on: CRAFT Brain-score concept integrated gradients CRP
Closest papers by meaning
- Decoupling Semantic Similarity from Spatial Alignment for Neural Networks. (NeurIPS 2024 · similarity 0.88)
- Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity (ICLR 2025 · similarity 0.88)
- Representational Difference Explanations (NeurIPS 2025 · similarity 0.88)
- When Does Closeness in Distribution Imply Representational Similarity? An Identifiability Perspective (NeurIPS 2025 · similarity 0.87)
- ReSi: A Comprehensive Benchmark for Representational Similarity Measures (ICLR 2025 · similarity 0.87)
- Characterizing Universal Object Representations Across Vision Models (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.