Research map / Label-efficient and robust learning / Recognition and long-tail
How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability
Zijian Wang, Yadan Luo, Liang Zheng, Zi Huang, Mahsa Baktashmotlagh
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 |
|---|---|---|---|
| Use neural-collapse-based NCTI instead of SFDA for ranking supervised models | Transferability estimation · 11 target datasets, supervised models | average weighted Kendall tau: 0.663 -> 0.726 | Improved |
| Use NCTI instead of SFDA for ranking self-supervised models | Transferability estimation · 11 target datasets, self-supervised models | average weighted Kendall tau: 0.708 -> 0.719 | Improved |
| Use NCTI instead of SFDA in terms of compute time | Transferability estimation · 11 datasets, running time | running time: approximately 10x speed up | Improved |
| Add the nearest-centroid term Sncc to the combination (ablation) | NCTI · Supervised models | average weighted Kendall tau | Mixed |
| Add the nearest-centroid term Sncc to the combination (ablation) | NCTI · Supervised and self-supervised combined | average weighted Kendall tau: +4.2% absolute | Improved |
| Combine Sseli with Svc instead of using either alone (ablation) | NCTI · Supervised and self-supervised models | average weighted Kendall tau | Improved |
| Use equal weights of 1 for the three terms (ablation) | NCTI · 11 datasets, weights 0.1 to 2 | average tau: tau >= 0.7 for most weight settings | No clear effect |
| Reduce target data to few samples per class (ablation) | NCTI · SUN397 | weighted Kendall tau | Mixed |
| Compute Svc on source features of pre-training data (ablation) | NCTI components · Source subset | weighted Kendall tau: 0.725 (supervised), 0.598 (self-supervised) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- LogME — LogME: Practical Assessment of Pre-trained Models for Transfer Learning (ICML 2021) · better on 2 of 2 shared results
- LEEP — LEEP: A New Measure to Evaluate Transferability of Learned Representations (ICML 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- The Impact of Geometric Complexity on Neural Collapse in Transfer Learning (NeurIPS 2024 · similarity 0.89)
- Neural Collapse Beyond the Unconstrained Features Model: Landscape, Dynamics, and Generalization in the Mean-Field Regime (ICML 2025 · similarity 0.87)
- Understanding the Transferability of Representations via Task-Relatedness (NeurIPS 2024 · similarity 0.86)
- Features are fate: a theory of transfer learning in high-dimensional regression (ICML 2025 · similarity 0.86)
- Implicit Modeling for Transferability Estimation of Vision Foundation Models (NeurIPS 2025 · similarity 0.86)
- Wide Neural Networks Trained with Weight Decay Provably Exhibit Neural Collapse (ICLR 2025 · similarity 0.85)
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.
