On the Reproducibility of Neural Network Predictions
This paper addresses the reproducibility challenge in machine learning for researchers and practitioners by identifying causes of prediction churn and proposing methods to mitigate it, which is an incremental improvement.
The authors investigate the problem of "churn" in neural networks, where independently trained models produce different predictions for the same input due to training randomness. They demonstrate churn on CIFAR and ImageNet and propose two mitigation strategies: minimum entropy regularizers and a novel co-distillation variant, both of which reduce churn while improving model accuracy.
Standard training techniques for neural networks involve multiple sources of randomness, e.g., initialization, mini-batch ordering and in some cases data augmentation. Given that neural networks are heavily over-parameterized in practice, such randomness can cause {\em churn} -- for the same input, disagreements between predictions of the two models independently trained by the same algorithm, contributing to the `reproducibility challenges' in modern machine learning. In this paper, we study this problem of churn, identify factors that cause it, and propose two simple means of mitigating it. We first demonstrate that churn is indeed an issue, even for standard image classification tasks (CIFAR and ImageNet), and study the role of the different sources of training randomness that cause churn. By analyzing the relationship between churn and prediction confidences, we pursue an approach with two components for churn reduction. First, we propose using \emph{minimum entropy regularizers} to increase prediction confidences. Second, \changes{we present a novel variant of co-distillation approach~\citep{anil2018large} to increase model agreement and reduce churn}. We present empirical results showing the effectiveness of both techniques in reducing churn while improving the accuracy of the underlying model.