word2vec. Dropout.

[willsmithorg]

Two of the most important papers in the history of ML, in my opinion, are:

word2vec - really the start of word embeddings. I don't know, but I guess that word embeddings are as useful in written Chinese as in a phonetic language like English. https://arxiv.org/abs/1301.3781

dropout - essential in neural networks, we now take this for granted. Sure, it's not SOTA now, but it was at the time, and it's nice to know the history of ideas. https://jmlr.org/papers/v15/srivastava14a.html