Multiple Instance Learning

Multiple Instance Learning (MIL) treats an example as a bag of instances and learns to predict from the bag when the informative regions or moments are unknown. It is useful for tasks with labels for whole slides, scans, videos, or documents but few or no labels for individual patches, slices, frames, or passages. Methods range from mean and max pooling to attention, transformers, graphs, and state-space models, often combined with pretrained features, spatial or temporal structure, prototypes, contrastive learning, and pseudo-labels. A central tension is that strong bag-level predictions do not guarantee faithful instance-level evidence, and more complex aggregators do not always outperform simple pooling.