Robust Self-Supervised Multi-Instance Learning with Structure Awareness

Multi-instance learning (MIL) is a supervised where each example labeled bag with many instances. The typical MIL strategies are to train an instance-level feature extractor followed by aggregating instances features as bag-level representation information. However, such highly depends on large number of datasets, which difficult get in real-world scenarios. In this paper, we make the first attempt propose robust Self-supervised Multi-Instance LEarning architecture Structure awareness (SMILEs) that learns unsupervised representation. Our proposed approach is: 1) permutation invariant order bag; 2) structure-aware encode topological structures among instances; and 3) against noise or permutation. Specifically, yield model without label information, augment multi-instance encoder maximize correspondence between representations same its different augmented forms. Moreover, capture from nearby bags, our framework optimal graph for bags these graphs optimized together message passing layers ordered weighted averaging operator towards contrastive loss. main theorem characterizes invariance Compared state-of-the-art baselines, SMILEs achieves average improvement 4.9%, 4.4% classification accuracy 5 benchmark datasets 20 newsgroups respectively. addition, show input corruption.