In modern society, communication systems play an essential role in everyday life. Up to this point, digital signal processing (DSP) has supported communications systems (radio, cellular, Internet, etc.) as well as the connected devices. With ever increasing demands on the amount, speed, reliability, and security of transmitted information, the underlying communications technologies need constant improvement. Artificial Intelligence (AI) has been tasked to replace DSP in order to meet forecasted performance demands, expand capabilities of current systems, as well as inspire future communications technologies. However, there are shortcomings of modern AI when trying to integrate into communications systems: interpretability ("Can we understand why modern AI methods succeed and fail?"), deployability ("Can modern AI operate on everyday electronics, without the need for the cloud?"), and generalizability ("Can these systems perform well in unforeseen circumstances?"). This thesis addresses technical problems withholding the future of communications with respect to these three shortcomings: (1) enabling complex-valued computations in real-valued neuarl networks (interpretability & generalizability), (2) novel neural network pruning methods enabling edge-ready AI (interpretability & deployability), and (3) a gray-box, physics-based AI simulation tool for fiber optic networks (interpretability & generalizability).