At the same time, a number of “pseudopatterns” will be generated in the final-storage
memory. In other words, random input will be fed into the network. Activation will spread
through both the early-processing and final-storage areas of the network. Crucially, the
output of this random input sent through the final-storage part of the network will be used
as a teacher for the early-processing part of the network. In other words, the earlyprocessing
memory will then learn a series of pseudopatterns produced in the final-storage
memory that reflect the patterns previously stored in final-storage. So, rather than
interleaving the real, originally learned patterns with the new input coming to the earlyprocessing
memory, we do the next best thing — namely, we interleave pseudopatterns that
are approximations of the previously stored patterns. Once the new pattern and the
pseudopatterns are learned in the early-processing area, the weights from the earlyprocessing
network are copied to the corresponding weights in the final-storage network