Autoencoders

The study of psychiatric and neurologic disorders typically involves the acquisition of a wide range of different types of data, such as brain images, electronic health records, and mobile phone sensors data. Each type of data has its unique temporal and spatial characteristics, and the process of extracting useful information from them can be very challenging. Autoencoders are neural networks that can automatically learn useful features and representations from the data; this makes them an ideal technique for simplifying the process of feature engineering in machine learning studies. In addition, autoencoders can be used for dimensionality reduction, denoising data, generative modeling, and even pretraining deep learning neural networks. In this chapter, we present the fundamental concepts of autoencoders and provide an overview of how they execute these tasks. Finally, we show some exemplary applications from brain disorders research.