SCALABLE MULTI-NEIGHBORHOOD LEARNING FOR CONVOLUTIONAL NETWORKS

In this paper we explore the role of scale for improved fea-ture learning in convolutional networks. We propose multi-neighborhood convolutional networks, designed to learn im-age features at different levels of detail. Utilizing nonlinear scale-space models, the proposed multi-neighborhood model can effectively capture fine-scale image characteristics (i.e., appearance) using a small-size neighborhood, while coarse-scale image structures (i.e., shape) are detected through a larger neighborhood. In addition, we introduce a scalable learning method for the proposed multi-neighborhood archi-tecture and show how one can use an already-trained single-scale network to extract image features at multiple levels of detail. The experimental results demonstrate the superior performance of the proposed multi-scale multi-neighborhood models over their single-scale counterparts without an in-crease in training cost. Index Terms — convolutional neural networks, scale-space models, feature extraction, image classification 1