Hi,
I have done some work in face recognition in the past, with neural networks. There are a couple of very promising new papers that may give the accuracy that you require but no implementation so far. The best one is by facebook research and it will take some time to implement it, but it basically is what powers (or used to power some iterations ago) facebooks face recognition.
You mentioned that you have a database of images which consist of faces side by side, how large is it? Also you say that you need at least 90% accuracy on the test set, which in my experience is a very bad estimate of how a model performs. (it is very easy to overfit something to get 90% on test :) )
Another kind of big question I have is where is this going to be used. You need a matlab implementation which is okay, but I would not recommend using any matlab in production, so maybe we should think about some alternatives? (tensorflow, pytorch?)
Hope we will work together!