News for hackers. Stuff that matters.
Re:NNCP: Lossless Data Compression with Neural Networks (2019) by labawi on Saturday May 22, 2021 @09:33AM (#27246359)
What do you mean by collisions?

As described, arithmetic coding does a 1:1 mapping of input-output, while NN deterministically generates probability distribution for the arithmetic coding. Conceptually, it would work with any deterministic / reproducible function, even a PRF or a running hash, though the point of NN is to give good distribution estimates for interesting inputs, thus short output from the arithmetic encoder.

Presumably, there are some limited choices during coding, which then need to be encoded as well.

There are no comments on this comment yet.