Abstract

CNN-based models currently provide state-of-the-art performance in image categorization tasks. While these methods are powerful in terms of representational capacity, they are generally not conceived with explicit means to control complexity. This might lead to scenarios where resources are used in a non-optimal manner, increasing the number of unspecialized or repeated neurons, and overfitting to data. In this work we propose CompactNets, a new approach to visual recognition that learns a hierarchy of shared, discriminative, specialized, and compact representations. CompactNets naturally capture the notion of compositional compactness, a characterization of complexity in compositional models, consisting on using the smallest number of patterns to build a suitable visual representation. We employ a structural regularizer with group-sparse terms in the objective function, that induces on each layer, an efficient and effective use of elements from the layer below. In particular, this allows groups of top-level features to be specialized based on category information. We evaluate CompactNets on the ILSVRC12 dataset, obtaining compact representations and competitive performance, using an order of magnitude less parameters than common CNN-based approaches. We show that CompactNets are able to outperform other group-sparse-based approaches, in terms of performance and compactness. Finally, transfer-learning experiments on small-scale datasets demonstrate high generalization power, providing remarkable categorization performance with respect to alternative approaches.