Deep Learning with R Cookbook

There's more...

Autoencoders can learn internal representations from raw input data. One of the challenges with these autoencoders is that they can be too specialized for the training data, that is, they overfit and do not generalize for new data. Regularization makes an autoencoder less sensitive to the input but, at the same time, minimizing reconstruction errors forces it to remain sensitive to capture more variations. Applying penalties to the loss function appropriately makes the model more robust and learn generalized features.

In this section, we will learn about two types of regularized autoencoders:

• Contractive autoencoders
• Sparse autoencoders

Contractive autoencoders: These are regularized autoencoders where a penalty is applied ...