MIMS (Multiple Instance Learning for Missing Annotations)
A learning framework that trains on labeled groups of instances when exact annotations are unavailable.
MIMS is a specialized adaptation of Multiple Instance Learning (MIL) designed for scenarios where precise per-instance labels are unavailable, but labels for groups of instances — called bags — can be obtained. Rather than requiring annotators to identify exactly which element within a collection is responsible for a given label, MIMS allows a model to learn from the weaker signal that at least one instance in a positive bag belongs to the target class. This makes it especially valuable in domains like medical imaging, where a radiologist might confirm that a scan contains a tumor without pinpointing its exact location, or in document classification, where a document is labeled by topic without sentence-level annotation.
The learning mechanism in MIMS operates by treating each bag as a training unit. A bag is labeled positive if it contains at least one instance satisfying the concept of interest, and negative if none of its instances do. The model must then learn a decision boundary at the instance level that is consistent with these bag-level constraints. This is typically framed as a constrained optimization problem, and various approaches — including support vector machines adapted for MIL, expectation-maximization methods, and deep neural network architectures — have been applied to solve it effectively.
The practical significance of MIMS lies in its ability to reduce the annotation burden without sacrificing model utility. Labeling individual instances in large datasets is often prohibitively expensive or requires domain expertise that is difficult to scale. By accepting coarser supervision, MIMS enables practitioners to train useful models on data that would otherwise be too costly to annotate fully. This aligns well with the growing emphasis on weakly supervised and semi-supervised learning paradigms in modern machine learning.
MIMS and related MIL approaches have found application across a wide range of fields, including drug activity prediction, remote sensing, natural language processing, and video analysis. As datasets grow larger and annotation budgets remain constrained, techniques that extract maximum signal from imperfect labels continue to gain relevance, positioning MIMS as a practical tool in the weakly supervised learning toolkit.
