Learning Augmentation from Data: A Case Study in Scientific Diagnostic Simulations

Deep neural networks are a popular machine learning technique for analyzing and extracting useful information from data. However, neural networks are usually overparameterized and require large amounts of example training data to avoid overfitting and maximize generalizability. For image classification applications, data augmentation has shown great success as a means of increasing the effective size of datasets. However, automated data augmentation for scientific applications has not been thoroughly explored. Scientific datasets would especially benefit from data augmentation since simulations and experimental data can be prohibitively expensive. In this paper, we explore the idea of population-based augmentation (PBA) to learn on scientific datasets. PBA involves multiple neural networks, each with different data augmentation schedules and policies, that are trained simultaneously and evaluated at intervals to find the best augmentation policy. We provide experimental results from an automated analysis of proton beam imaging energy spectrometer data (PROBIES).