PaperReading20200324

Normalizing flow models

• Motivation: VAE is good at extracting features but finds it hard to get probability for data points. For autoregressive models, you can get probability easily while you can not learn features. Here we have Normalizing flow models.
• Method: The key idea is to find a simple distribution which is easy to sample and get density then we use NN to project this distribution to a normalized density, from which we get the name"Normalizing". This NN function can be learnt and invertible so we can use the inverse function to get the distribution.
• Contribution: Another interesting generative model!