In praise of the autoencoder


When you consider all the machine learning (ML) algorithms, you'll find there is a subset of very pragmatic ones: neural networks. They usually require no statistical hypothesis and no specific data preparation except for normalization. The power of each network lies in its architecture, its activation functions, its regularization terms, plus a few other features. When you consider architectures for neural networks, there is a very versatile one that can serve a variety of purposes -- two in particular: detection of unknown unexpected events and dimensionality reduction of the input space. This neural network is called autoencoder.