Having worked on compression algos, any NN is just way to slow for (de-)compression. A potential usage of them is for coarse prior estimation in something like rANS, but even then the overhead cost would need to carefully weighted against something like Markov chains since the relative cost is just so large.