Distortion Metric

Distortion metrics quantify the difference between an original data source (e.g., image, audio, or model parameters) and its compressed or transformed representation, aiming to objectively measure information loss. Current research focuses on developing metrics that better align with human perception or downstream task performance, moving beyond traditional metrics like PSNR and exploring alternatives such as Wasserstein distance and f-divergences, often integrated into optimization schemes like alternating minimization or dual gradient ascent. These advancements are crucial for improving the efficiency and effectiveness of various applications, including image and video compression, robust reinforcement learning, and federated learning, by enabling more accurate assessment of data fidelity and enabling the development of more perceptually-aligned algorithms.