Researchers have developed GLOBE (Gradient Local-Balanced Extraction), a novel framework for coreset selection in on-device deep neural network training. GLOBE addresses limitations of existing methods by capturing evolving optimization dynamics and handling correlated samples through trajectory-aligned gradient matching. The framework formulates sample selection as a globally optimized sparse weighting problem, utilizing multiple training checkpoints to construct gradient trajectories and employing regularization techniques to induce sparsity and stabilize weights. Experiments show GLOBE outperforms existing methods in downstream test accuracy, especially at low data retention ratios. AI
IMPACT Enhances data efficiency in on-device AI training by improving coreset selection methods.
RANK_REASON The cluster contains a research paper detailing a new framework for coreset selection in machine learning. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
