In multiple instance learning, an object is represented by a set (bag) of feature vectors (instances) instead of a single feature vector or instance, as in traditional classification problems. For example, an image can be described by a bag of segments, where each segment is represented by a feature vector. Other examples include text documents, spam/regular email, molecule activity, structure of proteins.
Labels are given only for bags, but it is often assumed that instances have labels as well, and that these instance labels somehow influence the bag label. The standard assumption is “a bag is positive if and only if at least one of the instances is positive”. For example, for an image which is tagged as “tiger”, this means that at least one of the segments contains a tiger.
The challenge is to classify new bags using labels of bags from the training set. One approach is to try to infer which instances are actually positive, and label unseen bags accordingly. Another approach is to learn on bag-level, without finding out the underlying cause of the bag label.