Enhancing Normalizing Flow Performance with MLP-Mixer

Abstract

Recently, it has received a lot of attention due to the development of generative models. Normalizing flow (NF), one of the generative models, is a model that predicts the distribution by using a reversible function that makes it from a simple distribution to a complex distribution. NF is widely used in various fields and has the advantage of being able to obtain an accurate likelihood. However, NF has structural limitations due to its characteristic of being reversible, and unlike local information, it also tends not to fully utilize semantic information. In this study, we propose a new method to make more use of semantic information by using the mixer layer of MLP-Mixer, and the mixer layer suggests a method of reversibly changing it. The improvement of NF was shown through the method proposed in this study.