Information-Bottleneck Approach to Salient Region Discovery

We propose a new method for learning image attention masks in semi-supervised setting based on the Information Bottleneck principle. Provided with set of labeled images, mask generation model is minimizing mutual information between input and masked while maximizing same label. In contrast other approaches, our produces Boolean rather than continuous mask, entirely concealing masked-out pixels. Using synthetic datasets MNIST CIFAR10 SVHN datasets, we demonstrate that can successfully attend to features known define class.