Conceptual
Browse Instruct Curate

Variational Autoencoders in Deep Learning

A class of deep generative models that learn a probabilistic latent-variable representation of data by jointly training an encoder that maps inputs to a distribution over latent codes and a decoder that reconstructs inputs from sampled codes, optimizing a variational lower bound on the data likelihood. Students learn how VAEs generate new samples resembling the training data and how the reparameterization trick makes the model trainable by gradient descent.