A Appendix

The packages used in this work and their respective licenses are listed below: 1. PGDL Competition Starter Kit [1]; Apache 2.0 2. PyTorch [37]; BSD 3. PyTorch Lightning [38]; Apache 2.0 4. Tensorflow [36]; Apache 2.0 A.2 Calculating Conditional Mutual Information scores Given it's importance to our analyses, we reproduce the calculation for Conditional Mutual Information (CMI) scores presented in [1] here. Inter-class interpolation is performed similarly. The first set of plots shows the timing results for all datasets (Figure 4). Figure 5 shows the Gi-score CMI sensitivity to number of batches (mean and std. Pal-scores are oppositely correlated (anti-correlated). Mixup, and also include combinations of all Gi and Pal-scores ("PCA all Gi & Pal"), all Gi-scores V arious combinations do better at different tasks.