Abstract

The use of machine learning (ML)in the health sciences has exploded in recentyears. Unfortunately, ML products have not produced consistently reproducibleresults (McDermott et al. 2021). One major reason for reproducibility issues inhealth ML is that data are often incomplete. We looked at issues in the perfor-mance of a deep convolutional neural network on an Alzheimer’s Disease brainimaging data-set. We investigated if regression improves a deep CNN architec-ture’s accuracy when there are hidden categories. Surprisingly, we found thatregression drastically harmed our model’s generalizability. Additionally, we ex-plored the loss-landscape associated with various parameterizations of a base CNNarchitecture. We find that models with a higher number of parameters are moreprone to overfitting because their loss landscapes are more hilly