This advanced chapter has shown you one of the most interesting and simpler models that is able to generate data from a learned distribution using the configuration of an autoencoder and by applying variational Bayes principles leading to a VAE. We looked at the pieces of the model itself and explained them in terms of input data from the Cleveland dataset. Then, we generated data from the learned parametric distribution, showing that VAEs can easily be used for this purpose. To prove the robustness of VAEs on shallow and deep configurations, we implemented a model over the MNIST dataset. The experiment proved that deeper architectures produce well-defined regions of data distributions as opposed to fuzzy groups in shallow architectures; however, both shallow and deep models are particularly good for the task of learning representations.
By this point, you should feel confident in identifying the pieces of a VAE and being able to tell the main differences between a traditional...
