vae.md

VAE (2013)

• Problem: latent variable-based generation model
• Solution: propose a variational auto-encoder model with elbo loss and reparameterization tricks.
• Key method:
  • Variational lower bound: This is theoretical proof of why minimizing elbo loss is lower or equivalent to minimizing p(x), and also elbo loss enables us to sample z from certain distribution.
  • Reparametrization tricks: Modification from sampling to adding noise for backpropagation purposes.

Reference

@article{kingma2013auto,
  title={Auto-encoding variational bayes},
  author={Kingma, Diederik P and Welling, Max},
  journal={arXiv preprint arXiv:1312.6114},
  year={2013}
}