Multiple-Instance Learning

Multiple-Instance Learning (MIL) is a form of weak supervision. Instead of learning from examples labeled one by one, the model receives groups of examples called bags, and it knows only the label of the whole bag, not that of each instance it contains. It answers a very concrete question: how do you train a classifier when you know that «something» in a collection of data matters, but not which of its pieces?

That asymmetry reshapes the problem. In classic supervised learning every instance carries its own label; in MIL the label sits one level up, and the model must work out which instances justify it without anyone pointing them out.

The standard assumption

The classic formulation rests on a simple premise, which James Foulds and Eibe Frank named the «standard assumption» in their 2010 review: a bag is positive if it contains at least one positive instance and negative only if all of its instances are negative. A single guilty piece condemns the whole bag; every piece must be innocent to acquit it. That asymmetry gives MIL its character and its difficulty, because the instance responsible for the label hides among many irrelevant ones. Later work has proposed variants that relax the premise —requiring a proportion or a combination of instances rather than a single one— useful when the label depends on the collection as a whole and not on one element.

Where it comes from and what it is for

The term was coined by Thomas Dietterich, Richard Lathrop and Tomás Lozano-Pérez in 1997, in a paper in the journal Artificial Intelligence on drug-activity prediction. Their case was molecular: a single molecule adopts many three-dimensional shapes, or conformations, and it is active if at least one of them fits the target receptor. The chemist knows whether the molecule works, but not which conformation is responsible. They modeled each molecule as a bag and each conformation as an instance, and solved it with axis-parallel rectangles. The payoff is direct: labeling bags is far cheaper than labeling every instance. Noting that «this biopsy contains cancer» is trivial next to marking every tumor cell, and MIL squeezes that coarse label to spare the fine-grained manual work.

Current uses

Today the liveliest example is computational pathology. A whole-slide image holds billions of pixels; it is cut into thousands of patches, and that set is the bag. From the clinical record we know only whether the patient has cancer —the bag label— with no one having flagged the malignant patches. The system by Gabriel Campanella and colleagues (Nature Medicine, 2019) trained clinical-grade classifiers this way from tens of thousands of images, while proposals such as the attention-based MIL of Maximilian Ilse and colleagues (2018) and the CLAM method of Ming Lu and colleagues (2021) added mechanisms that also point to which patches drove the decision. Beyond medicine, MIL supports weakly supervised object detection —you know an object is in the image, not where— and document classification, where the label belongs to the full text rather than to each sentence.