Hands-On Mathematics for Deep Learning

Real-valued non-volume preserving

So far in this chapter, we have covered two very popular generative neural network architectures—VAEs and GANs—both of which are quite powerful and have brought about tremendous results in generating new data. However, both of these architectures also have their challenges. Flow-based generative models, on the other hand, while not as popular, do have their merits.

Some of the advantages of flow-based generative models are as follows:

• They have exact latent-variable inference and log-likelihood evaluation, whereas in VAEs, we can only approximately infer from latent variables, and GANs cannot infer the latent as they do not have an encoder.
• They are efficient to parallelize for both synthesis and inference. ...