How to train a deep or shallow network embedded into another system?

I need to embed a deep or shallow network into another algorithm to do certain jobs. How can I calculate the loss function and pass it to the training session so that the output of the outer system is optimized compared to pre-determined performance?