Well, I know that there have been some experiments with pattern recognition (fuzzy logic, neural networks) where they fed it 'hits' and then charged it to produce one. Kinda like when they did that with pictures of beautiful people and then had it make a picture- in both cases the result was rather bland, but not too far off the mark.
The idea that a trained neural network could compare input to it's 'perfect' template and then calculates the overlap is viable, but probably not practical (yet). That seems to be what they're offering from what I've read, and a comparison with the DB of the neural network would be very quick (The dissection of the input data should however take a while depending on the width of parameters). The accuracy of the result depends on how well the network was trained, how many aspects of the input are compared and human-made adjustments for exceptions, training errors and such.
So, yes, it could work to a degree. I'd file this particular site under 'toy' though, because I have yet to see a neural network that was trained so well that it could tackle a complex beast like the effect of music on the human brain*, mainly because we still haven't figured out how exactly our brain works. There were some interesting efforts in combining genetic algorithms (evolution based self-writing networks) with AI last I checked on the topic, and complex tasks in areas that are explained well (ie we know most of the system parameters) can be solved. For example, a few years ago a trained network was able to perform the landing operation of a jet.
*) See for example the yet unsuccessful attempts of using neural networks as a translation tool. Language is comparable to music in width of parameters and complexity
Edit:Fixed some Germanglish
Last edited by µB (Feb 12, 2010 9:11 pm)