Neural Preprocessing Methods

When using data, real or artificial, it is often convenient to perform on them some form of preprocessing. This preprocessing is often aimed at extracting the most important features of the data, thus performing dimensionality reduction, or at changing the data to a more convenient format (e.g. decorrelating the components of the data vectors). This paper presents an overview of unsupervised, neural-style preprocessing algorithms. It starts by analyzing criteria for unsupervised feature extraction, and then proceeds to describe some of the most important feature extraction algorithms. Here, the emphasis is placed on linear algorithms that extract principal components or their linear combinations.