Other point to be considered is the training stage. Basically, the rule given by equation (13) updates the parameters of the system. This is much more closer to a (classical) neural network approach than the expression (4).
Despite of these advantages, a doubt about this QNN is that its neural network approach is virtual in the sense that it is just a biased interpretation of an approximated quantum model of the system.
Although this argument may be consistent, it does not discard the model because a neural network should have three basic elements: (1) An operator to process the input signal(s) (equation (7) in the above model); (2) A test to decide if the results is the desired one (expression (11)); (3) A rule to adapt parameters if need (equation (12)). In the above model, these relations are not virtual ones in the sense that the final result is a system that reproduce a desired behavior (a gate, for example).