2019/05/30 by Aditya Grover, Christopher G. Chute, Grover, Aditya +7 · 1 citation
Computer Science · #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Neural and Evolutionary Computing (cs.NE) #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1905.12892
openalex publication_date 2019/05/30 · openalex created_date 2022/07/29 · openalex updated_date 2026/07/28
Given datasets from multiple domains, a key challenge is to efficiently\nexploit these data sources for modeling a target domain. Variants of this\nproblem have been studied in many contexts, such as cross-domain translation\nand domain adaptation. We propose AlignFlow, a generative modeling framework\nthat models each domain via a normalizing flow. The use of normalizing flows\nallows for a) flexibility in specifying learning objectives via adversarial\ntraining, maximum likelihood estimation, or a hybrid of the two methods; and b)\nlearning and exact inference of a shared representation in the latent space of\nthe generative model. We derive a uniform set of conditions under which\nAlignFlow is marginally-consistent for the different learning objectives.\nFurthermore, we show that AlignFlow guarantees exact cycle consistency in\nmapping datapoints from a source domain to target and back to the source\ndomain. Empirically, AlignFlow outperforms relevant baselines on image-to-image\ntranslation and unsupervised domain adaptation and can be used to\nsimultaneously interpolate across the various domains using the learned\nrepresentation.\n