One of the criticisms of the pseudo-recurrent architecture as shown in Figure 1 is that the
mechanism of copying the weights from the early-processing sub-network to the final-storage
sub-network lacks plausibility. In particular, this constraint also means that both sub-networks
have to have identical sizes and connection topologies. In the final experiment, then, this
constraint is relaxed and information is transferred from one sub-network to the other
exclusively by means of pseudopatterns. Both the Ebbinghaus savings measure and the exact
recognition measure of forgetting were used to demonstrate that catastrophic forgetting is
significantly reduced in this type of network, just as it is in the original system that relies on
weight-copying. In other words, the weight-copying mechanism, while certainly more
efficient than transferring information by pseudopatterns, is not a necessary feature of the
pseudo-recurrent memory model. This is important because it is clearly more difficult to
justify weight-copying in a cognitive model that the transfer of information using the
pseudopattern mechanism described in this paper