A Additional Results In addition to C = 0 and λ

CLIP described in 2, we train two more instantiations of it by keeping either of the two consistency regularizers active in the loss objective (Eq. CLIP as only cross-modal consistency regularizer term is added to the loss objective. CLIP on most of the experiments discussed in the main text to understand their zero-shot transfer ability on standard datasets and robustness to natural distribution shifts. A.1 Zero-shot Transfer Table 7 presents our results of the zero-shot transfer experiment described in 3.1. CLIP outperforms its sub-variants and the CLIP model on the ImageNet1K dataset.