Clip_Dataset__NeurIPS2022_ (10)

We visualize samples of the class "broom" from the reference We find that the data efficiency (i.e., how fast the error The two models' logit predictions are ensembled with equal weights Output mixing results for two CLIP models trained on YFCC-3M + CC-3M mixture and RedCaps-3M respectively . Ensemble outputs of CLIPs trained on different data sources and dataset sizes (red and orange lines), taken from the same stage of training (i.e., epoch), lie on the linear trend of training a We provide proofs of main theoretical claims in Section 6. F .1 Proof of Theorem 1