Impact of Trivial Solutions on Classification Tasks Using CNNs

This paper investigated the impact on the training process of convolutional neural networks (CNN) caused by unconscious perturbations during data collection in the form of unintentional encoding the ground truth label. For this purpose, three possible scenarios (background, dead pixel and illumination gradient) were created and their influence on the classification accuracy were evaluated using the BloodMNIST dataset. In practice, especially in sensitive areas such as medical contexts, the generation of appropriately large and diverse datasets for training CNNs is often challenging due to limited access, lack of financial resources, but also due to a lack of awareness of the need for diverse data. This contribution showed that transfer learning was less vulnerable to label encoded perturbations than models learned from scratch, even when scratch models performed better on the collected dataset. Simulation results support the assumption, that transfer learning provides a better choice to address challenging data acquisition.