Autoencoders can practically learn very interesting data projections that can help to reduce the dimensionality of the data without much data loss in the lower dimensional space. The encoder compresses the input and selects the most important features, also known as latent features, during compression. The decoder is the opposite of encoder, and it tries to recreate the original input as closely as possible. While encoding the original input data, autoencoders try to capture the maximum variance of the data using lesser features. 

In this recipe, we will build a deep autoencoder to extract low dimensional latent features and demonstrate how we can use this lower-dimensional feature set to solve various learning problems such as regression, classification, and more. Dimensionality reduction decreases training time significantly. While reducing the dimensionality, autoencoders also learn non-linear features present in the data, hence enhancing the model's performance.