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

This paper introduces a new ensemble-based approach to reduce the data and computation costs of accurate classification. When faced with a new test case, a low cost classifier is used first, only moving to a higher cost approach if the initial classifier does not have a high degree of confidence in its projection. This multi-stage strategy can be used with any set of classifiers and does not require additional training. The approach is first applied to reduce the amount of data required to classify test images; it is found to be effective for problems in which at least some fraction of cases can be correctly classified based upon coarser data than are typically used. For neural networks performing digit recognition, for example, the proposed approach reduces the number of bytes of data read by 60% to 85% with less than 5% reduction in accuracy. For the ImageNet data, the number of bytes read by the typical network is reduced by 20% with less than 5% reduction in accuracy -- and in some cases, the resource savings reach 40%. The second application is to reduce computational complexity, with simpler neural networks used for test cases that are easier to classify and complex networks used for more difficult cases. For classification both of digits and of ImageNet images, computation cost is reduced by as much as 82% to 89% with less than 5% reduction in accuracy. The results also show that, for situations in which computational cost is not a concern, calculating multiple models' projections and selecting the one from the most confident classifier can increase classification accuracy on ImageNet by as much as two percent over the best standalone classifier considered here.