Reproduce “A Framework and Benchmark for Deep Batch Active Learning for Regression” 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
David Holzmüller, Viktor Zaverkin, Johannes Kästner, Ingo Steinwart
Researchers introduced a modular framework and a new open-source benchmark for Deep Batch Active Learning (BMDAL) in regression, featuring a novel LCMD-TP selection method combined with sketched Neural Tangent Kernels. This approach demonstrated superior sample efficiency, consistently reducing RMSE compared to existing methods across 15 diverse tabular datasets.
Key insights
- Network-dependent kernels, particularly those leveraging neural network gradients (k_grad), are crucial for effective regression BMDAL, providing a more comprehensive representation than network-independent or last-layer features.
- Applying sketching to high-dimensional kernels like k_grad significantly reduces computational runtime without sacrificing accuracy, making these methods scalable for large datasets.
- The LCMD-TP selection method, combined with sketched k_grad, offers a superior balance of informativeness and diversity, leading to enhanced sample efficiency in deep regression tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.