infrastructurefactneutralPre-training with Masked Image Modeling is inherently more robust to heterogeneous data than Contrastive Learning in distributed self-supervised learningMachine Learning27 Jul 2026http://arxiv.org/abs/2607.02447v1