Paper on efficient data pruning accepted at DMLR Workshop at ICML 2023

The paper "Repeated Random Sampling for Minimizing the Time-to-Accuracy of Learning“ by Patrik Okanovic, Roger Waleffe, Vasilis Mageirakos, Konstantinos E. Nikolakakis, Amin Karbasi, Dionysis Kalogerias, Nezihe Merve Gürel and Theodoros Rekatsinas has been accepted at DMLR Workshop at ICML 2023. It is based on Patrik Okanovic’s master thesis at ETH Zürich. The paper proposes using Repeated Sampling of Random Subsets (RS2) to achieve faster time-to-accuracy during training. It finds RS2 as being the best option when compared against thirty state-of-the-art methods for data pruning and data distillation across four datasets.