multiple instance learning

Gold definitionUpdated Apr 2, 2026

Definition

Multiple Instance Learning (MIL) is a supervised learning paradigm where labels are assigned to "bags" of instances, not individual instances. It's commonly used when individual instance labels are unavailable or ambiguous, such as in medical image analysis, to infer bag properties from critical instances.

At a glance

Executive summary

Multiple Instance Learning (MIL) is a machine learning method for situations where data is grouped into "bags" and only the bag has a label, not individual items within it. It's particularly useful in medical imaging, like analyzing whole-slide pathology images, to identify disease based on small, critical regions without needing to label every single cell.

TL;DR

Multiple Instance Learning helps computers learn from groups of data where only the group has a label, often by finding the few important items within the group that determine its label.

Key points

Learns from labels assigned to bags of instances, inferring bag properties from constituent instances, often assuming one positive instance makes a bag positive.
Addresses the challenge of scarce or impossible instance-level annotations, enabling analysis of complex data like whole-slide images.
Used by researchers and engineers in medical imaging (histopathology), drug discovery, computer vision, and bioinformatics.
Unlike traditional supervised learning requiring instance-level labels, MIL operates with only bag-level labels, reducing annotation burden.
Increasing focus on interpretability and evidence-efficiency in MIL, as exemplified by ReaMIL's ability to identify minimal sufficient evidence sets.

Use cases

Digital Pathology: Diagnosing diseases (e.g., cancer subtypes like LUAD vs. LUSC, IDC vs. Others) from whole-slide histopathology images by identifying critical regions.
Drug Discovery: Predicting the toxicity or efficacy of chemical compounds based on the properties of their constituent molecules.
Image Retrieval: Finding images relevant to a query where an image might contain multiple objects, and only one needs to match.
Remote Sensing: Classifying land cover types from satellite imagery where a pixel block (bag) might contain various terrain features.

Also known as

MIL, Multi-instance learning, Bag-level learning, Weakly supervised learning (often related context)