Leveraging Reproduction-Error Representations for Multi-Instance Classification

Multi-instance learning deals with the problem of classifying bags of instances, when only the labels of the bags are known for learning, and the instances themselves have no labels. In this work, we propose a method that trains autoencoders for the instances in each class, and recodes each instance into a representation that captures the reproduction error for this instance. The idea behind this approach is that an autoencoder trained on only instances of a single class is unable to reproduce examples from another class properly, which is then reflected in the encoding. The transformed instances are then piped into a propositional classifier that decides the latent instance label. In a second classification layer, the bag label is decided based on the output of the propositional classifier on all the instances in the bag. We show that this reproduction-error encoding creates an advantage compared to the classification of non-encoded data, and that further research into this direction could be beneficial for the cause of multi-instance learning.