Reproduce “Influence Guided Sampling for Domain Adaptation of Text Retrievers” 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
Meet Doshi, Vishwajeet Kumar, Yulong Li, Jaydeep Sen
An adaptive sampling framework, Inf-DDS, dynamically reweighs training datasets for dense text retrievers using influence-based reward signals. This approach, developed by IBM Research AI, improves domain adaptation and retrieval performance across various benchmarks while offering computational efficiency and stable sampling.
Key insights
- Influence-based reward signals offer a more stable and reliable mechanism for adaptive data sampling than noisy gradient estimates, providing a direct measure of a dataset's impact on target performance.
- Significant computational efficiencies in dynamic sampling can be gained by reusing intermediate computations and judiciously subsampling when updating the sampling distribution's parameters.
- Optimal sampling for dense retrievers is critical for domain adaptation, enabling effective prioritization of diverse datasets (e.g., low-resource languages) without negatively affecting performance on dominant ones.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.