One limitation of the scheme so far is that the quantum system really only implements the forward phase of the network. The back-propagation of error, or learning phase, must be calculated off-line, after which the slits are altered accordingly. It might be better if the learning phase were implemented directly in the quantum system as well, by having the desired interference patterns and actual interference patterns directly cause the changes in the weight slits.
Imagine a setup similar to the one already described, but with the following additions. Behind the plate, there is another particle beam, directed back toward the original barrier. Furthermore, there are slits in the plate, which will allow the second beam to pass through and hit the original barrier, which has a photo-sensitive plate mounted on its back. Thus, the second particle beam will cause interference patterns on the back side of the original barrier.
The goal would be to have the setup work like this: the plate itself would calculate the difference between the actual and desired interference patterns (perhaps by having something like the negative of the target pattern projected onto the plate), then this could cause certain slits to open in the plate, causing characteristic interference patterns on the back of the original barrier. These patterns would in turn cause the weight slits in the barrier to move according to whatever learning rule is being employed.
Of course, implementation-dependent speculations such as these may be premature, or irrelevant, since the principal reason for using both the barrier/slit/plate setup and back-propagating feed-forward networks was not for ease of implementation, but ease of explication.