Fast and Accurate Inference with Adaptive Ensemble Prediction for Deep Neural Networks

Ensembling multiple predictions is a widely-used technique to improve the accuracy of various machine learning tasks. One obvious drawback of the ensembling is its higher execution cost during inference. In this paper, we first describe our insights on relationship between the probability of the prediction and the effect of ensembling with current deep neural networks; ensembling does not help mispredictions for inputs predicted with a high probability even when there is a non-negligible number of mispredicted inputs. This finding motivates us to develop a new technique called adaptive ensemble prediction, which achieves the benefits of ensembling with much smaller additional execution costs. 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 cost significantly while achieving similar accuracy to the naive ensembling. We also showed that our statistically rigorous confidence-level-based termination condition reduces the burden of the task-dependent parameter tuning compared to the naive termination based on the pre-defined threshold in addition to yielding a better accuracy with the same cost.