Deep Gaussian mixture model based instance relevance estimation for multiple instance learning applications

Multiple instance learning (MIL) is a type of supervised learning, where instead of receiving a collection of individually labeled examples, the learner is given weakly labeled bags of instances. If the bag contains at least one positive instance, the bag is assigned a positive label, otherwise, the bag is assigned a negative label. The positive bags in MIL may contain instances from different classes, which results in instance-level ambiguity in the bag and complicates the learning process. In this case, identifying relevant instances is important in bag classification and model interpretation. To identify the relevant instances in MIL, this paper proposes a deep subspace-based Gaussian mixture model instance relevance estimation network with Fisher vector encoding (DGMIR-FV). To be specific, the proposed approach uses an estimation network for instance relevance estimation and selects the instances from each bag based on relevance scores. Afterwards, selected instances are encoded using the Fisher vector encoding and fed to an ensemble network for classification. Compared to the existing MIL pooling methods and encoding schemes, the DGMIR-FV improves the model's generalization ability by employing estimation network for instance relevance estimation and incorporating relevant instances in the encoding process. The experimental results demonstrate the efficiency of DGMIR-FV on several MIL benchmark datasets.