2019/01/24 by Isabela Albuquerque, Albuquerque, Isabela, João Monteiro +9 · 3 citations
Computer Science · Engineering · #Advanced Numerical Analysis Techniques #Anomaly Detection Techniques and Applications #FOS: Computer and information sciences #Generative Adversarial Networks and Image Synthesis #Machine Learning (cs.LG) #Machine Learning (stat.ML)
paper · pdf · doi:10.48550/arxiv.1901.08680
openalex publication_date 2019/01/24 · openalex created_date 2022/07/30 · openalex updated_date 2026/07/28
Recent literature has demonstrated promising results for training Generative\nAdversarial Networks by employing a set of discriminators, in contrast to the\ntraditional game involving one generator against a single adversary. Such\nmethods perform single-objective optimization on some simple consolidation of\nthe losses, e.g. an arithmetic average. In this work, we revisit the\nmultiple-discriminator setting by framing the simultaneous minimization of\nlosses provided by different models as a multi-objective optimization problem.\nSpecifically, we evaluate the performance of multiple gradient descent and the\nhypervolume maximization algorithm on a number of different datasets. Moreover,\nwe argue that the previously proposed methods and hypervolume maximization can\nall be seen as variations of multiple gradient descent in which the update\ndirection can be computed efficiently. Our results indicate that hypervolume\nmaximization presents a better compromise between sample quality and\ncomputational cost than previous methods.\n