|
The best-correlating
candidate (here a source network) is installed into the network, its input
weights are frozen (indicated by the solid arrow), and training reverts back
to output phase, in which all weights (indicated by the dashed arrows)
entering the output units are trained in order to reduce error.
|