| 摘要 |
A method for the supervised teaching of a recurrent neutral network (RNN) is disclosed. A typical embodiment of the method utilizes a large (50 units or more), randomly initialized RNN with a globally stable dynamics. During the training period, the output units of this RNN are teacher-forced to follow t he desired output signal. During this period, activations from all hidden units are recorded. At the end of the teaching period, these recorded data are use d as input for a method which computes new weights of those connections that feed into the output units. The method is distinguished from existing traini ng methods for RNNs through the following characteristics: (1) Only the weights of connections to output units are changed by learning - existing methods fo r teaching recurrent networks adjust all network weights. (2) The internal dynamics of large networks are used as a "reservoir" of dynamical components which are not changed, but only newly combined by the learning procedure - existing methods use small networks, whose internal dynamics are themselves competely re-shaped through learning.
|
