SetFlow: A Generative Architecture for Multiple Instance Learning in Representation Space

SetFlow introduces a novel generative architecture designed to address data scarcity and weak supervision in machine learning applications like mammography, where Multiple Instance Learning (MIL) is commonly used. The model operates directly in representation space, modeling entire MIL bags as sets rather than individual instances. It combines flow matching with a Set Transformer-inspired design to handle permutation-invariant inputs while capturing intra-bag dependencies between instances. Conditioned on both class labels and input scale, SetFlow generates coherent and semantically structured sets of representations. This approach overcomes limitations of existing methods that fail to capture interactions within bags. The work is documented in arXiv preprint 2604.16362v1, categorized as a cross announcement. By generating structured sets, SetFlow aims to enhance performance in real-world applications where traditional augmentation methods fall short.