Dataset Pruning using Evolutionary Optimization

Abstract

Data is key to training deep neural networks. A common demand for individual data units is their abundance and diversity. However, it is barely investigated what is actually an informative data unit and how the amount of data relates to the neural network performance. In this study, we utilize evolutionary algorithms to optimize data usage during deep neural network training. We test multiple medical classification and segmentation datasets as being key tasks in medical imaging and found that this so-called dataset pruning removes rather unimportant data elements. Depending on how much we punished the incorporation of data, we found that across tasks and datasets, a critical amount of data is incorporated by the algorithm itself. This shows that future research not only needs to incorporate abundant data but rather relevant data.