Research map / Label-efficient and robust learning / Recognition and long-tail
Improving neural network representations using human similarity judgments
Lukas Muttenthaler, Lorenz Linhardt, Jonas Dippel, Robert A. Vandermeulen, Katherine Hermann, Andrew Kyle Lampinen, Simon Kornblith
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 |
|---|---|---|---|
| apply gLocal transform | CLIP-ViT-L/14 (WIT) · Entity-13 5-shot | 5-shot accuracy: 65.34 -> 71.94 | Improved |
| apply gLocal transform | CLIP-ViT-L/14 (WIT) · CIFAR100 anomaly detection | AUROC: 91.41 -> 97.19 | Improved |
| apply gLocal transform | CLIP-ViT-L/14 (LAION-2B) · DTD 5-shot | 5-shot accuracy: 69.49 -> 68.44 | Hurt |
| apply gLocal transform | ImageNet models (AlexNet) · CIFAR100 5-shot | 5-shot accuracy: 28.70 -> 23.29 | Hurt |
| apply gLocal transform | ResNet-50 (ImageNet) · Entity-13 5-shot | 5-shot accuracy: 49.27 -> 59.18 | Improved |
| add local contrastive loss and shrinkage (gLocal) to naive alignment (ablation) | CLIP-ViT-L/14 (WIT) · 5-shot FS (avg.) | 5-shot accuracy: 47.62 -> 69.91 | Improved |
| add local contrastive loss and shrinkage (gLocal) to naive alignment (ablation) | CLIP-ViT-L/14 (WIT) · ImageNet nearest neighbors | closest neighbor preserved (%): 12.2 -> 76.3 | Improved |
| add local loss to global transform (ablation) | CLIP ViT-L · ImageNet nearest neighbors | closest neighbor preserved (%): 73.7 -> 76.3 | Improved |
| gLocal instead of naive transform on human alignment (ablation) | CLIP-ViT-L/14 (WIT) · THINGS odd-one-out | odd-one-out accuracy (%): 60.13 -> 60.05 | No clear effect |
| apply gLocal on top of Tip-Adapter | CLIP-ViT-L/14 (WIT) · CIFAR100 5-shot | 5-shot accuracy: 45.79 -> 64.51 | Improved |
| train on human-adversarial triplets (ablation) | gLocal transform · downstream tasks | accuracy | Hurt |
Closest papers by meaning
- Evaluating alignment between humans and neural network representations in image-based learning tasks (NeurIPS 2024 · similarity 0.88)
- Discriminating image representations with principal distortions (ICLR 2025 · similarity 0.88)
- Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity (ICLR 2025 · similarity 0.87)
- Characterizing Universal Object Representations Across Vision Models (NeurIPS 2026 · similarity 0.87)
- When does perceptual alignment benefit vision representations? (NeurIPS 2024 · similarity 0.87)
- Representational Similarity via Interpretable Visual Concepts (ICLR 2025 · similarity 0.86)
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.
