Hands-On Unsupervised Learning with Python

Denoising autoencoders

A very helpful application of autoencoders is not strictly related to their ability to find lower-dimensional representations, but relies on the transformation process from input to output. In particular, let's assume a zero-centered dataset, X, and a noisy version whose samples have the following structure:

In this case, the goal of the autoencoder is to remove the noisy term and recover the original sample, x_i. From a mathematical viewpoint, there are no particular differences between standard and denoising autoencoders; however, it's important to consider the capacity needs for such models. As they have to recover ...