Reproduce “Information Guided Regularization for Fine-tuning Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mandar Sharma, Nikhil Muralidhar, Shengzhe Xu, Raquib Bin Yousuf, Naren Ramakrishnan
Researchers from Virginia Tech and Stevens Institute of Technology introduced an information-guided regularization method called guided dropout for fine-tuning large language models. This approach assigns layer-wise dropout probabilities based on Fisher information, leading to improved generalization and consistently better performance on 9 out of 10 GLUE benchmark tasks, particularly in data-paucity settings, without adding computational overhead.
Key insights
- Fisher information reliably approximates the Hessian matrix in LMs, revealing the loss landscape geometry and parameter sensitivity.
- An exceedingly small fraction of parameters and specific layers within LMs possess significantly high Fisher scores, disproportionately influencing the loss landscape and generalization ability.
- Perturbing highly sensitive parameters leads to sharp minimizers (poor generalization), while perturbing less sensitive ones maintains wide minimizers (better generalization).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.