Recurrent Neural Networks with Delays

Recurrent neural networks (RNN) have been introduced for the modelization of non linear dynamic systems. In spite of their potential ability to implement systems of arbitrary complexity, they are often avoided because of large training times. For many problems, local recurrent networks or time delay feed forward architectures perform well enough. They are thus often preferred to more general systems, even though they have limited capabilities for handling sequences and can only build short term memory of past events [Frasconi et al. 92].