I think it is a cool 'tangible' demo of how non-human readable data remains learnable for ML purposes. I've recently tried using Nilsimsa and sdeep hashes to re

I think it is a cool 'tangible' demo of how non-human readable data remains learnable for ML purposes. I've recently tried using Nilsimsa and sdeep hashes to reduce the input dimensionality of natural text, with some loss in accuracy on similarity tasks. A combination of byte-pair encoding and local-sensitive hashing might prove a more stable combination. Throw some gzip function that way and we may be able to reduce input corpus file sizes immensely.