Teacher Forcing

In all other types of recurrent neural networks, any connection from an output node to some other node in the network will communicate the last value computed by this output node. If there was an error in this value, it gets propagated to the rest of the network. In teacher forcing, any network that receives an input from an output node will receive the desired output value, as opposed to the one computed by the network. Williams and Zipser (1988) report that using this method greatly reduced the learning time for some of their experiments, but that it is completely unnecessary for others.

Teacher Forcing

In all other types of recurrent neural networks, any connection from an output node to some other node in the network will communicate the last value computed by this output node. If there was an error in this value, it gets propagated to the rest of the network. In teacher forcing, any network that receives an input from an output node will receive the desired output value, as opposed to the one computed by the network.

Williams and Zipser (1988) report that using this method greatly reduced the learning time for some of their experiments, but that it is completely unnecessary for others.

Back to the Table of Content

Back to the Previous subject

To the Next Subject