Pushing the barriers or independent validation on the latest research is not a good place to start if you're new to the area.
Go and implement an image classifier, tweak the parameters and the topology and see how the results and training changes.
Switch to a time series model / text model, and learn the difference between convolution / recurrent networks.
Start playing with non-sequential topologies, custom objective functions, Q-Learning,
Once you have a grasp of these basics go back and read the papers, and you'll find that you understand them a lot more, and you'll see where they're pushing the boundaries.
