Multiple Instance Learning (MIL) is a specific type of weakly supervised learning problem concerned with training samples as a set of instances, known as a bag. Labeling is declared for the entire bag and not for instances. MIL handles problems with incomplete knowledge of labels in training datasets. The primary goal of the MIL is based on the labeled bags as training data, and it classifies the unseen instances.
The significance of MIL is dealing with weakly annotated data that reduces the annotation cost. The methods of MIL are divided into instance space and bag space methods based on the reasoning space. Characteristics of MIL problems are task or prediction level(instance level vs. bag level), bag composition, data distributions, and label ambiguity. Some of the learning algorithms of MIL are Learning Axis-Parallel Concepts, Diverse Density (DD) and its EM version, Expectation-Maximization version of Diverse Density (EM-DD), Citation kNN, Support Vector Machine for multi-instance learning, Multiple-decision tree, and MIL with the neural network.
Application fields of MIL are biology, chemistry, computer vision, image, audio sound and document classification, web mining, time series, medical imaging, and Reinforcement learning with MIL. Some of the specific applications of MIL are sentiment analysis in text, computer-aided diagnosis, drug activity predictions, object localization in images, content-based image retrieval, and molecular classification. Future investigations of MIL methods deal with huge amounts of data, multiple modalities, and class imbalance.