Autoencoders in generative modeling, feature extraction, regression, and classification

Autoencoders are a family of neural network algorithms with a wide range of applications. Potential uses include but are not limited to feature extraction, sampling, denoising, dimensionality reduction, and generative modeling. Generally speaking, an autoencoder consists of two parts. An Encoder function tries to copy input to output, and the Decoder attempts to reconstruct the input from the output. But, the point is that copying does not take place completely. Forcing some restrictions on encoder and decoder functions, they tend to extract just necessary and salient features. Autoencoders are robust and data-friendly, so they are suitable for big datasets. Autoencoders have been used successfully as generative models for de novo molecular design, chemometric data analysis, robust classification on molecular biology datasets, dimension reduction mechanism for various imaging datasets, etc. This chapter introduces mathematical formulations of well-known autoencoders, training procedures, and some state-of-the-art uses.