Vanishing Contributions: A Unified Approach to Smoothly Transition Neural Models into Compressed Form