ConceptKL vanishing / KL散度消失 / posterior collapse
Posterior Collapse
A common failure mode in VAE training in which the posterior distribution of latent variables degenerates to the prior, causing the latent variables to carry no meaningful information and reducing the model to a plain autoencoder. It is typically mitigated through techniques such as KL annealing or KL weighting.
Timeline (last 90 days)
Sep 11
The "posterior collapse" (KL vanishing) problem in VAE training can render the latent variable ineffective, requiring techniques such as annealing and weighting to ensure the latent variable carries meaningful information.
Unverified50%