What characterises posterior collapse in a VAE?

Question: Which observation most directly indicates posterior collapse in a variational autoencoder? A) The reconstruction loss approaches zero while the approximate posterior becomes deterministic B) The aggregate posterior differs substantially from the prior, although individual posteriors remain broad C) qφ(z|x) becomes close to p(z) for most x, so the decoder can generate outputs while using little information from z D) The decoder likelihood becomes multimodal and the reparameterisation gradient has high variance Correct: C Explanation: During posterior collapse, the inferred latent distribution carries little input-specific information because it closely matches the prior. The decoder effectively ignores the latent variable, and the KL term may approach zero even though reconstruction quality can remain reasonable. Topic: advanced ML / VAE / posterior collapse