Supplementary Material for Neural Complexity Measures A Proof of Motivating Bound We first invoke the following lemma which relates the empirical and true cumulative distribution

Steps 1 2 4 8 16 No regularization 4.17 4.04 4.05 4.04 4.05 L In Figure B.1, we show additional visualizations of regression tasks. Figure B.2 shows additional experiments, where we additionally compare against stronger baselines Figure B.3 shows additional visualizations of loss surfaces, and reveals that the NC-regularized loss Circles represent the targets and plus signs represent predictions. Best viewed zoomed in. 3 set of m We train NC with batch size bs and the Adam optimizer with learning rate lr. C.2 Classification T ask Learner We provide further details about the single-task experiments. Using such splits, we trained NC as usual.