Informative Imitation

Beguš's term for how a speech GAN learns: the generator never sees training data; it must produce audio that fools a discriminator and encode a code (binary in fiwGAN, one-hot in ciwGAN) that a Q-network can recover from the output. The requirement to be maximally informative forces it to encode linguistically meaningful distinctions (lexical and sublexical) rather than arbitrary acoustics. The paper contrasts this with autoencoder replication and transformer prediction: imitation is a distinct learning regime, analogous to a child learning to encode meaning into imitated sounds, and is the mechanism behind imagitation and spontaneous concatenation. Whether imitation is really non-predictive (vs. the corpus's predictive-intelligence consensus) is an open tension.