Dataset Pruning

Dataset pruning studies how to train effectively on a smaller subset of examples, reducing compute and sometimes limiting bias or environmental cost. Methods rank examples by loss, influence, or training dynamics, or select for coverage and diversity using geometric, graph-based, class-aware, and semantic criteria; label-free and pseudo-label approaches help when annotations are scarce. Evaluations span image classification benchmarks such as CIFAR and ImageNet, as well as retrieval, forecasting, scientific data, and domain-specific tasks. A central tension is that useful subsets must balance redundancy reduction with retaining hard, rare, or representative examples, especially when label noise and dependencies between samples make generic selection rules unreliable.