• If you have a target that is hard to predict, make an alternative targets (auxiliary objective) where this is the “hardness to predict label”
    • this is a pesudolabel that you generate
    • how to make this column:
        1. trained some baseline DNN to predict the original target
        1. for the predictions, label it as follows: (y_true = 1 and y_pred < 0.2) or (y_true = 0 and y_pred > 0.8) label 1