Deep learning methods that involve classification and prediction models using data that has a response or a dependent variable are part of supervised deep learning methods. When working with structured or unstructured data, there are situations where the response variable is either not available or not used. Applications of deep learning networks that do not use a response variable are classified as unsupervised deep learning methods. For example, an application of deep learning may involve image data from which we want to extract important features in order to achieve dimension reduction. Another example involves handwritten images that contain unwanted noise and a deep network is used for denoising the images. In such situations, autoencoder networks have been found to be very useful for performing unsupervised deep learning tasks.

Autoencoder neural networks make use of an encoder and decoder network. When the image data is passed through an encoder and the resulting dimension is lower than that of the original image, the network is forced to extract only the most important features from the input data. And then the decoder part of the network reconstructs the original data from whatever is available from the output of the encoder. In Chapter 6, Applying Autoencoder Neural Networks Using Keras, we provide an illustration of applying an autoencoder neural network for dimension reduction, de-noising, and image correction when working with image data using R.