|Topic:||Symmetric learning for Variational Autoencoders (VAE)|
|Supervisor:||doc. Boris Flach, Dr. rer. nat. habil.|
|Announce as:||Diplomová práce, Semestrální projekt|
|Description:||The standard ELBO based learning approach for VAEs requires to fix the prior distribution p(z) in the latent space in closed form. This prevents its use in cases when this distribution is accessible by sampling only. In view of our long term goal to combine VAEs with discriminative learning of the encoder, we proposed a novel, symmetric algorithm for learning VAEs. It has the advantage that both the data distribution and the latent noise distribution can be accessible by sampling only. Moreover it generalises to hierarchical VAEs leading to a layer-wise learning algorithm for them.
The goal of this project is threefold:
(1) Conduct a thorough experimental comparison of the ELBO based learning and the novel symmetric learning approach. This should be done for deep hierarchical VAEs on a large dataset (CelebA or similar).
(2) Analyse theoretical properties of the novel learning algorithm (likelihood bounds, consistency)
(3) Conduct first steps towards combining generative and discriminative learning for hierarchical VAEs (interpreting the decoder layers as feedback connections of the encoder)
|Bibliography:||Diederik P. Kingma, Max Welling, An Introduction to Variational Autoencoders, arXiv:1906.02691|