Autoencoder neural networks consist of two main parts:

• The first part is called the encoder, which reduces the dimensions of the input data. Generally, this is an image. When data from an input image is passed through a network that leads to a lower dimension, the network is forced to extract only the most important features of the input data.
• The second part of the autoencoder is called the decoder and it tries to reconstruct the original data from whatever is available from the output of the encoder. The autoencoder network is trained by specifying what output this network should try to match.

Let's consider some examples where we will use image data. If the output that's specified is the same image that was given as input, then after training, the autoencoder network is expected to provide an image with a lower resolution that retains the key features of the input image but misses some finer details that were part of the original input image. This type of autoencoder can be used for dimension reduction applications. Since autoencoders are based on neural networks that are able to capture non-linearity in data, they have superior performance compared to methods that only use linear functions. The following diagram shows the encoder and decoder parts of autoencoder networks:

If we train autoencoders so that the input image contains some noise or non-clarity and the output as the same image but without any noise, then we can create denoising autoencoders. Similarly, if we train autoencoders with such input/output images where we have images with and without glasses, or with and without a mustache, and so on, we can create networks that help with image correction/modification.

Next, we will look at three separate examples of how to use an autoencoder: using dimension reduction, image denoising, and image correction. We will start by using autoencoders for dimension reduction.