Rate distortion Multiple Instance Learning for image classification

In this paper, we model image classification as a Multiple Instance Learning (MIL) problem, by regarding each image as a bag composed of different regions/patches (i.e., instances). Motivated by the fact that a bag is determined by the most positive instance or the least negative instance, which is also called “witness”, we propose a new algorithm to take advantage of witnesses to improve the performance of MIL for image classification task. In the frame of Rate Distortion (RD), we regard MIL as a source coding of each instance to witness or itself, and then the distortion function is measured by the loss of the discriminant model trained on these encoded instances. Hence compared with the existing algorithms, our proposed RDMIL algorithm has the following advantages. First, the probabilistic approach in source coding well illustrates the generative process of witnesses and measures the different importance of instances. Second, the discriminant model trained in a large-margin approach sufficiently considers the diverse influences from the instances, and thus has a strong discriminative ability. The resulted objective function is decomposed into two convex sub-problems, and we especially design a sequential method to effectively optimize the RD sub-problem. Experimental results on two real-world datasets demonstrate the proposed algorithm is effective and promising.