it truly is shown that The straightforward pre-training activity of predicting which caption goes with which impression is really an economical and scalable way to understand SOTA impression representations from scratch https://k2spiceshop.com/product/liquid-k2-on-paper-online/