Fast and Accurate Inference with Adaptive Ensemble Prediction in Image Classification with Deep Neural Networks

Ensembling multiple predictions is a widely used technique to improve the accuracy of various machine learning tasks. In image classification tasks, for example, averaging the predictions for multiple patches extracted from the input image significantly improves accuracy. Using multiple networks trained independently to make predictions improves accuracy further. One obvious drawback of the ensembling technique is its higher execution cost during inference. If we average 100 predictions, the execution cost will be 100 times as high as the cost without the ensemble. This higher cost limits the real-world use of ensembling, even though using it is almost the norm to win image classification competitions. In this paper, we describe a new technique called adaptive ensemble prediction, which achieves the benefits of ensembling with much smaller additional execution costs. Our observation behind this technique is that many easy-to-predict inputs do not require ensembling. Hence we calculate the confidence level of the prediction for each input on the basis of the probability of the predicted label, i.e. the outputs from the softmax, during the ensembling computation. If the prediction for an input reaches a high enough probability on the basis of the confidence level, we stop ensembling for this input to avoid wasting computation power. We evaluated the adaptive ensembling by using various datasets and showed that it reduces the computation time significantly while achieving similar accuracy to the naive ensembling.