Problem-dependent attention and effort in neural networks with an application to image resolution

This paper assesses a new classification approach that examines low-resolution images first, only moving to higher resolution images if the classification from the initial pass does not have a high degree of confidence. This multi-stage strategy for classification can be used with any classifier and does not require additional training. The approach is tested on five common datasets using four different classification approaches. It is found to be effective for cases in which at least some fraction of cases can be correctly classified using coarser data than are typically used. neural networks performing digit recognition, for instance, the proposed approach reduces the resource cost of classifying test cases by 60% to 85% with less than 5% reduction in accuracy.