Towards Faithful Data Selection: Statistical Limits and Scalable Algorithms
This work proposes a swap-based local search that decreases the bound at every accepted step and can refine the output of any selector without worsening its bound, and provides experimental results on CIFAR10 and CIFAR100 showing that it matches existing data selection baselines.
Ahmed Mehdi Inane, Mahdi Haghifam, G. Dziugaite et al.
· 0 citations