Un-mix: Rethinking Image Mixtures for Unsupervised Visual Representation Learning

The recently advanced unsupervised learning approaches use the siamese-like framework to compare two "views" from same image for representations. Making views distinctive is a core guarantee that methods can learn meaningful information. However, such frameworks are sometimes fragile on overfitting if augmentations used generating not strong enough, causing over-confident issue training data. This drawback hinders model subtle variance and fine-grained To address this, in this work we aim involve soft distance concept label space contrastive-based task let be aware of degree similarity between positive or negative pairs through mixing input data space, further collaboratively loss spaces. Despite its conceptual simplicity, show empirically with solution -- Unsupervised mixtures (Un-Mix), subtler, more robust generalized representations transformed corresponding new space. Extensive experiments conducted CIFAR-10, CIFAR-100, STL-10, Tiny ImageNet standard ImageNet-1K popular SimCLR, BYOL, MoCo V1&V2, SwAV, etc. Our proposed mixture assignment strategy obtain consistent improvement by 1~3% following exactly hyperparameters procedures base methods. Code publicly available at https://github.com/szq0214/Un-Mix.