Learning prototypes for multiple instance learning

Abstract

Multiple instance learning (MIL) is a weakly supervised learning method that works on the labeled bag of instances data. A prototypical network is a popular embedding approach in MIL. They overcome the common problems that other MIL approaches may have to deal with including dimensionality, loss of instance-level information, and complexity. They demonstrate competitive performance in classification. This work proposes a simple model that provides a permutation invariant prototype generator from a given MIL data set. We aim to find out prototypes in the feature space to map the collection of instances (i.e. bags) to a distance feature space and simultaneously learn a linear classifier for MIL. Another advantage of prototypical networks is that they are commonly used in the machine learning domain to facilitate interpretability. Our experiments on classical MIL benchmark data sets demonstrate that the proposed framework is an accurate and efficient classifier compared to the existing approaches.