Sounds like "unlearning" is really just "reduce the probability of sampling" from some latent "learned space" and quantizing reduces the efficacy of the slight change in sampling.