it is actually shown that the simple pre-training process of predicting which caption goes with which picture can be an economical and scalable way to know SOTA impression representations from scratch over a dataset of https://k2spiceshop.com/product/liquid-k2-on-paper-online/