Reproduce “GLOBE: Trajectory-Aligned Gradient Matching with Structured SparseOptimization for Coreset Selection” 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
Hetian Liu, Jin Cui, Mengcheng Shi, Yanbin Hu, Xinyue Long, Boran Zhao, Pengju Pen
GLOBE selects real training examples by matching multi-checkpoint gradient trajectories rather than relying on a single optimization snapshot. Combining first- and projected second-order statistics with Group LASSO, Elastic Net, and class-balanced selection, it achieved the best reported accuracy across all 18 dataset–retention settings, with particularly strong gains at 10% data retention.
Key insights
- Training-time importance is dynamic: replacing multi-checkpoint trajectories with only the final checkpoint reduced CIFAR-10 accuracy from 90.21% to 88.53% at 10% retention.
- Matching only first-order behavior is insufficient; projected second-order information and structured regularization help preserve gradient variability and reduce redundancy among correlated examples.
- The method shifts substantial work to coreset construction, since per-sample gradients, multiple checkpoints, projections, and clustering are required; runtime and memory costs are not fully reported.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.