DRIVE weights fine-tuning samples to limit forgetting without source data
Original title: DRIVE: Fine-tuning via Data Contribution- and Diversity-aware Weighting with Prior Regularization
Per-sample weights combine Shapley anti-forgetting and adaptation signals with diversity and a prior.
DRIVE is a fine-tuning weight-assignment method for when pre-training data is unavailable and catastrophic forgetting remains a problem. It assigns per-sample weights through a differentiable objective that splits Shapley contribution into anti-forgetting and adaptation terms, reshapes them with a covariance diversity term, and stabilizes them with a sample-characteristic prior. The authors claim stronger stability-plasticity guarantees, faster convergence, and better robustness than existing weighting schemes. Code is released on GitHub.
- The method targets single-stage fine-tuning without access to source pre-training data.
- Shapley additivity separates anti-forgetting from new-task adaptation.
- A covariance diversity term and a sample-characteristic prior reshape the weights.
- Authors report superior overall performance and stronger parameter-estimation robustness.
From the original article
1 min read at arxiv.orgSingle-stage fine-tuning is pervasive, yet catastrophic forgetting persists when the source pre-training data is inaccessible. Data weighting affords a natural lever to modulate sample influence.
By Qing Liu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.