Capacity Visual Attention Networks

Inspired by recent work in machine translation and object detection, we introduce an attentionbased model that automatically learns to extract information from an image by adaptively assigning its capacity across different portions of the input data and only processing the selected regions of different sizes at high resolution. This is achieved by combining two modules: an attention sub-network which uses a mechanism to model a human-like counting process and a capacity sub-network. This subnetwork efficiently identifies input regions for which the attention model output is most sensitive and to which we should devote more capacity and dynamically adapt the size of the region. We focus our evaluation on the Cluttered MNIST, SVHN, and Cluttered GTSRB image datasets. Our findings indicate that the proposed model is able to drastically reduce the number of computations, compared with traditional convolutional neural networks, while maintaining similar or better performance.