Unsupervised Models C2.1 Feedforward models

Feedforward unsupervised models cover a wide range of neural networks with various applications. In this section, we discuss three widely used models. (i) Kohonen’s selforganizing map, also called the Kohonen network , the self-organizing feature map, or the topological map, is intended to map a high-dimensional space into a oneor twodimensional space, preserving the topology of the input space; it has a strong biological plausibility and is basically intended to be used in applications where preserving the topology between input and output spaces is important (e.g. control, inverse mapping, image compression). It is an unsupervised model, but can be extended to a supervised one by adding a supplementary layer. In addition to the topology-conserving property, the Kohonen model also acts as a vector quantizer. (ii) The neural gas is another vector quantization algorithm that may be considered as a neural network method because it relies on the same principle of adaptation, may be represented in the form of a feedforward graph, and may be described by the same formalism as used in many other neural models. It is different from the Kohonen map in the sense that it does not have the topology preserving property but it generally performs better giving a smaller final distortion error. (iii) The neocognitron is a complex feedforward model formed by several layers each containing a large number of neurons. Its goal is to automatically detect features in two-dimensional arrays of points through self-organization and reinforcement principles. The network is built to be insensitive to shifts in position of the patterns or of small parts of them, thus also allowing for distorted patterns. The network is primarily intended to be used in feature extraction and pattern recognition tasks, for example in OCR (optical character recognition). C2.1.1 Kohonen’s self-organizing map C2.1.1.