it is actually shown that the simple pre-education undertaking of predicting which caption goes with which graphic is definitely an efficient and scalable way to discover SOTA image representations from scratch on the https://k2spiceshop.com/product/liquid-k2-on-paper-online/