Hi Rohit,
I have worked with neural networks in the past, and I guess what you are using is a simple feed forward network with a single perceptron at the output. I had faced a similar problem a while ago - i.e. the output of my RBFN was stunningly dissimilar to what was predicted by MATLAB. And guess what? I reverted to C. The best thing to do!
Bottom line - been there, done that, can do it again. Let me know if you'd like me to work on this. I sure would love to :P
Cheers,
Sushil