Hands-On Artificial Intelligence for Beginners

Decoder

Like the standard autoencoder, the decoder in the VAE is a backward convolutional network, or a deconvolutional network. In processing the decoding, data is sampled from the generation stochastically (randomly), making the VAE one of the few models that can directly sample a probability distribution without a Markov chain Monte Carlo method. As a result of the stochastic generation process, the encoding that we generate from each pass will be a different representation of the data, all while maintaining the same mean and standard deviation. This helps with the decoder's sampling technique; because all encodings are generated from the same distribution, the decoder learns that a latent data point and its surrounding points are all ...