This is fascinating - I always thought the symbolic representation offered by LSH or quantization is wonderfully efficient but in terms of accuracy always seemed to dissapoint (or at least when compared to floating point vectors) - but neural hashing, if it's accurate and efficient - that is cool