Exploring the dimensionality reduction component of unsupervised deep learning

Dimensionality reduction is a technique that can be useful in cases where a faster runtime is needed to train and perform inference on your model or when the model has a hard time learning from too much data. The most well-known unsupervised deep learning method for dimensionality reduction is based on autoencoders, which we discussed in Chapter 5, Understanding Autoencoders. A typical autoencoder network is trained to reproduce the input data as an unsupervised learning method. This is done through the encoder-decoder structure. At inference time, using only the encoder will allow you to perform dimensionality reduction as the outputs of the encoder will contain the most compact representation, which can fully reconstruct the original input data. Autoencoders can support different modalities, with one modality at any one time, which makes it a very versatile unsupervised dimensionality reduction method...