Multiple Instance Learning (MIL) ist ein spezielles Paradigma in maschinellem Lernen that addresses scenarios where labels are assigned to sets, or bags, of instances rather than to individual instances. This approach is particularly useful in situations where it is challenging or impractical to obtain labels for every single instance. For example, in medizinische Bildgebung, a bag might represent a set of images from a patient, and the label indicates whether the patient has a specific condition, while individual images may contain varying levels of information.
In MIL, a bag is considered positive if at least one instance within it is positive; otherwise, it is labeled as negative. This unique labeling structure leads to distinct learning algorithms designed to extract features and make predictions at the bag level while considering the individual instances’ contributions. Techniques such as bagging, instance selection, and attention mechanisms are often employed to improve the performance of MIL models.
One significant advantage of MIL is its ability to leverage weak supervision, which can reduce the costs associated with data labeling in various applications, including drug discovery, Bildklassifikation, and object detection. By enabling models to learn from aggregated information, MIL opens new avenues for effectively learning from complex datasets.