To elaborate on Noe's comment, the article is suggesting the use of LSH where the hashing function is learned by a neural network such that similar vectors correspond to similar hashes via Hamming weight (whilst enforcing some load factor). In effect, a good hash is generated by a neural network. It appears Elastiknn a prioi chooses the hash function? Not sure, not my area of knowledge.

This approach seems feasible tbh. For example, a stock's historical bids/asks probably don't deviate greatly from month to month. That said, the generation of a good hash is dependent on the stock ticker, and a human doesn't have the time to find a good one for every stock at scale.