![[2505.11771] Residual Feature Integration is Sufficient to Prevent Negative Transfer](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2505.11771] Residual Feature Integration is Sufficient to Prevent Negative Transfer
Summary
This paper presents a novel approach to prevent negative transfer in transfer learning by integrating residual features from pretrained models with target-specific encoders, ensuring robust performance across various tasks.
Why It Matters
Negative transfer is a significant challenge in transfer learning, where using source data can hinder performance on target tasks. This research provides a theoretical foundation and practical solution to mitigate this issue, enhancing the reliability of machine learning models in diverse applications.
Key Takeaways
- Residual feature integration can effectively prevent negative transfer.
- The method offers theoretical guarantees for convergence rates comparable to training from scratch.
- Empirical results demonstrate robustness against distribution shifts and label noise.
- Supports multimodal adaptations in pretrained models.
- Advances the understanding of safe transfer learning practices.
Computer Science > Machine Learning arXiv:2505.11771 (cs) [Submitted on 17 May 2025 (v1), last revised 14 Feb 2026 (this version, v2)] Title:Residual Feature Integration is Sufficient to Prevent Negative Transfer Authors:Yichen Xu, Ryumei Nakada, Linjun Zhang, Lexin Li View a PDF of the paper titled Residual Feature Integration is Sufficient to Prevent Negative Transfer, by Yichen Xu and 3 other authors View PDF HTML (experimental) Abstract:Transfer learning has become a central paradigm in modern machine learning, yet it suffers from the long-standing problem of negative transfer, where leveraging source representations can harm rather than help performance on the target task. Although empirical remedies have been proposed, there remains little theoretical understanding of how to reliably avoid negative transfer. In this paper, we investigate a simple yet remarkably effective strategy: augmenting frozen, pretrained source-side features with a trainable target-side encoder that adapts target features to capture residual signals overlooked by models pretrained on the source data. We show this residual feature integration strategy is sufficient to provably prevent negative transfer, by establishing theoretical guarantees that it has no worse convergence rate than training from scratch under the informative class of target distributions up to logarithmic factors, and that the convergence rate can transition seamlessly from nonparametric to near-parametric when source represen...
