next up previous contents
Next: 4.1.1 Why abstract recurrent Up: 4. Abstract recurrent neural Previous: 4. Abstract recurrent neural


4.1 Motivation

First, favorable properties of recurrent neural networks (RNN) are mentioned, and reasons are given for using an abstract RNN that is based on the statistics of the training data rather than a network of neurons. Second, this section illustrates that a pattern completion might end in a local minimum if the completion is realized as a gradient descent in a potential field build on top of a mixture model. Thus, a different strategy is required.



Subsections

Heiko Hoffmann
2005-03-22