Attention-to-Embedding Framework for Multi-instance Learning

Abstract

AbstractWe present an attention-to-embedding framework that explicitly addresses the challenge posed by multi-instance learning (MIL) classification tasks, where learning objects are bags containing various numbers of instances. Two key issues of this work are to extract relevant information by determining the relationship between the bag and its instances, and to embed the bag into a new feature space. To respond to these problems, a network with the popular attention mechanism is designed that assigns a new representation and a class probability vector to a given instance in the bag. In addition, compared with the traditional MIL methods, we offer a new embedding function according to the assigned results of instances to process the bag embedding that is unrelated to the distance metric. As a result, MIL challenges will be reduced to single-instance learning (SIL) problems that can be solved using basic machine learning algorithms such as SVM. Extensive experiments on thirty-four data sets demonstrate that our proposed method has the best overall performance over other state-of-the-art MIL methods. This strategy, in particular, has a substantial advantage on web data sets and better stability. Source codes are available at https://github.com/InkiInki/AEMI.KeywordsAttentionEmbeddingMulti-instance learningNetwork