How to reduce both training and validation loss without causing overfitting or underfitting? : r/learnmachinelearning