A new unsupervised neural approach to stationary and non-stationary data

Dealing with time-varying high dimensional data is a big problem for real time pattern recognition. Non-stationary topological representation can be addressed in two ways, according to the application: life-long modeling or by forgetting the past. The G-EXIN neural network addresses this problem by using life-long learning. It uses an anisotropic convex polytope, which, models the shape of the neuron neighborhood, and employs a novel kind of edge, called bridge, which carries information on the extent of the distribution time change. In order to take into account the high dimensionality of data, a novel neural network, named GCCA, which embeds G-EXIN as the basic quantization tool, allows a real-time non-linear dimensionality reduction based on the Curvilinear Component Analysis. If, instead, a hierarchical tree is requested for the interpretation of data clustering, the new network GH-EXIN can be used. It uses G-EXIN for the clustering of each tree node dataset. This chapter illustrates the basic ideas of this family of neural networks and shows their performance by means of synthetic and real experiments.