A neural network is trained with too many parameters relative to the training set size. What is the primary risk?