Revisiting multiple instance neural networks. (February 2018)

• Record Type: Journal Article
• Authors: Wang, Xinggang; Yan, Yongluan; Tang, Peng; Bai, Xiang; Liu, Wenyu
• Is Part Of: Pattern recognition. Volume 74(2018:Feb.)
• Page Start: 15
• Page End: 24
• Publication Date: 2018-02
• Subjects: Multiple instance learning -- Neural networks -- Deep learning -- End-to-end learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.08.026
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We revisit the problem of solving MIL using neural networks (MINNs), which are ignored in current MIL research community. Our experiments show that MINNs are very effective and efficient. We proposed a novel MI-Net which is centered on learning bag representation in the neural networks in an end-to-end way. Recent deep learning tricks including dropout, deep supervision and residual connections are studied in MINNs. We find deep supervision and residual connections are effective for MIL. In the experiments, the proposed MINNs achieve state-of-the-art or competitive performance on several MIL benchmarks. Moreover, it is extremely fast for both testing and training, for example, it takes only 0.0003 s to predict a bag and a few seconds to train on MIL datasets on a moderate CPU. Abstract: Of late, neural networks and Multiple Instance Learning (MIL) are both attractive topics in the research areas related to Artificial Intelligence. Deep neural networks have achieved great successes in supervised learning problems, and MIL as a typical weakly-supervised learning method is effective for many applications in computer vision, biometrics, natural language processing, and so on. In this article, we revisit Multiple Instance Neural Networks (MINNs) that the neural networks aim at solving the MIL problems. The MINNs perform MIL in an end-to-end manner, which take bags with a various number of instances as input and directly output the labels of bags. All of the parameters in a MINN can be optimized via back-propagation. Besides revisiting the old MINNs, we propose a new type of MINN to learn bag representations, which is different from the existing MINNs that focus on estimating instance label. In addition, recent tricks developed in deep learning have been studied in MINNs; we find deep supervision is effective for learning better bag representations. In the experiments, the proposed MINNs achieve state-of-the-art or competitive performance on several MIL benchmarks. Moreover, it is extremely fast for both testing and training, for example, it takes only 0.0003 s to predict a bag and a few seconds to train on MIL datasets on a moderate CPU.