PaperReading20200323
CanChen
ggchen@mail.ustc.edu.cn
Recently I was too busy to update my blogs and today I am going to share something about generative models from CS236 courses.
VAE
- Motivation: Based on AE, some constraints are being put on the latent variable and used to change the distribution and make this distribution what we want. "variational" means changing latent distribution.
- Method: Just construct a lower bound for the data distribution and maximize this bound with optimization techs.
- Contribution: I found two tricks, REINFORCE and reparameterization, interesting. First, "REINFORCE" means that the reward function is in fact f(x) and the agent is in fact the data distribution. With a better reward, the gradient is correspondingly larger. Second, the interesting part of reparameterization is that we can train a network to model its distribution.