Hands-On Unsupervised Learning with Python

Variational autoencoders

Let's consider a dataset, X, drawn from a data generating process, p_data. A variational autoencoder is a generative model (based on the main concepts of a standard autoencoder), which was proposed by Kingma and Welling (in Auto-Encoding Variational Bayes, Kingma D. P. and Welling M., arXiv:1312.6114 [stat.ML]), aimed at reproducing the data-generating process. In order to achieve this goal, we need to start from a generic model based on a set of latent variables, z, and a set of learnable parameters, θ. Given a sample, x_i ∈ X, the probability of the model is p(x, z; θ). Hence, the goal of the training process is to find the optimal parameters that maximize the likelihood, p(x; θ), which can be obtained by marginalizing ...