Efficient Modeling of Contextual Mappings by Context-dependent Feedforward and Recurrent Neural Nets

The paper addresses the issue of solving real world problems by neural nets. The motivations for using contextual information in modeling complex dependencies between input data are discussed. The models of a context-dependent neuron and multi-layer perceptron are recalled along with a brief discussion on their properties and efficient training algorithms. Then the latest models of context-dependent radial basis function, recurrent, self-organizing and hybrid nets are introduced and provided with training algorithms. An example of application of context-dependent neural nets to gas load prediction is presented.