How many types of variational auto-encoders are there?

Question

I have been studying about auto-encoders and variational auto-encoders. I would like to know how many variants of VAEs are there today.

If there are many variants, can they be used for feature extraction for complex reinforcement learning tasks like self-driving cars?

Answer 1

There are many variations of the original VAE (proposed in the 2013 paper Auto-Encoding Variational Bayes), with different purposes (such as the generation of discrete data or graphs). Of course, I cannot enumerate all of them, so here I will list only the ones I am currently aware of.

• Conditional variational auto-encoder (CVAE) (2015, NeurIPS)
• Adversarial autoencoders (AAE) (2015, ICLR)
• Ladder variational autoencoder (LVAE) (2016, NeurIPS)
• VAE/GAN (2016, PMLR)
• Vector quantised variational autoencoder (VQ-VAE) (2017, NeurIPS)
• Denoising variational autoencoder (DVAE) (2017, AAAI)
• Grammar variational autoencoder (GVAE) (2017, ICML)
• Syntax-directed variational autoencoder (SD-VAE) (2017, ICLR)
• PixelVAE (2017)
• GraphVAE (2018, ICANN)
• Constrained graph variational autoencoder (CGVAE) (2018, NeurIPS)
• Character VAE (CVAE) (2018, ACS central science)
• Regularized VAE (RVAE) (2018, NeurIPS)
• Junction tree variational autoencoder (JT-VAE) (2018, PMLR)
• InfoVAE/MMD-VAE (2019, AAAI)
• Molecular hypergraph grammar variational autoencoder (MHG-VAE) (2019, ICML)
• DAG variational autoencoder (D-VAE) (2019, NeurIPS)
• Denoising adversarial autoencoders (iDAAE and DAAE) (2019, IEEE Transactions on Neural Networks and Learning Systems)
• MACRo-mIcro Disentangled Variational Auto-Encoder (MacridVAE) (2019, NeurIPS)

VAEs have been used for drug design (see e.g. JT-VAE) and in reinforcement learning (see e.g. world models). I don't know if they have been used for self-driving cars, but it's possible.