Research map / Detection and segmentation / CNN features
Duality Diagram Similarity: a generic framework for initialization selection in task transfer learning
Kshitij Dwivedi, Jiahui Huang, Radoslaw Martin Cichy, Gemma Roig
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 |
|---|---|---|---|
| add z-score normalization (Q, D) to feature comparison (ablation) | DDS (f=cosine) · Taskonomy winrate correlation | Spearman correlation with transfer ranking: 0.742 -> 0.864 | Improved |
| add z-score normalization (Q, D) to feature comparison (ablation) | DDS (f=linear kernel) · Taskonomy winrate correlation | Spearman correlation with transfer ranking: 0.632 -> 0.842 | Improved |
| add z-score normalization (Q, D) to feature comparison (ablation) | DDS (f=Laplacian) · Taskonomy winrate correlation | Spearman correlation with transfer ranking: 0.815 -> 0.860 | Improved |
| add z-score normalization (Q, D) to feature comparison (ablation) | DDS (f=Pearson) · Taskonomy winrate correlation | Spearman correlation with transfer ranking: 0.823 -> 0.856 | Improved |
| use layer or group normalization instead of z-score (ablation) | DDS (f=Laplacian) · Taskonomy winrate correlation | Spearman correlation with transfer ranking: 0.860 -> 0.806 | Hurt |
| add unbiased centering of similarity matrices (ablation) | DDS (f=Laplacian) · Taskonomy winrate correlation | Spearman correlation with transfer ranking: 0.691 -> 0.860 | Improved |
| add unbiased centering of similarity matrices (ablation) | DDS (f=euclidean) · Taskonomy winrate correlation | Spearman correlation with transfer ranking: 0.613 -> 0.850 | Improved |
| use DDS (cosine) instead of RSA for model selection | Taskonomy model selection · Taskonomy affinity correlation | Spearman correlation with transfer ranking: 0.777 -> 0.862 | Improved |
| use DDS instead of attribution-map methods (DeepLIFT) | Taskonomy model selection · Taskonomy affinity correlation | Spearman correlation with transfer ranking: 0.681 -> 0.862 | Improved |
| compute DDS from Pascal VOC images instead of Taskonomy images (ablation) | DDS (f=Laplacian) · Pascal VOC segmentation transfer ranking | Spearman correlation with transfer ranking: 0.5779 -> 0.765 | Improved |
| compute DDS from NYUv2 images instead of Pascal VOC images (ablation) | DDS (f=Laplacian) · Pascal VOC segmentation transfer ranking | Spearman correlation with transfer ranking: 0.765 -> 0.521 | Hurt |
| branch decoder from block 3 instead of block 1 (ablation) | ResNet-50 encoder · NYUv2 depth | log RMSE (lower better): 0.322 -> 0.207 | Improved |
Closest papers by meaning
- Understanding the Transferability of Representations via Task-Relatedness (NeurIPS 2024 · similarity 0.87)
- Features are fate: a theory of transfer learning in high-dimensional regression (ICML 2025 · similarity 0.86)
- How Not to Stitch Representations to Measure Similarity: Task Loss Matching Versus Direct Matching (AAAI 2025 · similarity 0.84)
- Representational Similarity via Interpretable Visual Concepts (ICLR 2025 · similarity 0.84)
- Knowledge Transfer from Vision Foundation Models for Efficient Training of Small Task-specific Models (ICML 2024 · similarity 0.83)
- Exploiting Task Relationships in Continual Learning via Transferability-Aware Task Embeddings (NeurIPS 2025 · similarity 0.83)
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.
