On Large Deviation Principles for Large Neural Networks