Hands-On Machine Learning for Algorithmic Trading

Fixing corrupted data with denoising autoencoders

The autoencoders we've discussed so far are designed to reproduce the input despite capacity constraints. An alternative approach trains autoencoders with corrupted input to output the desired, original data points.

Corrupted input are a different way of preventing the network from learning the identity function; instead, they extract the signal or salient features from the data. Denoising autoencoders have been shown to learn the data-generating process of the original data, and have become popular in generative modeling where the goal is to learn the probability distribution that gives rise to the input.