Analyzing and Improving Optimal-Transport-based Adversarial Networks
Jaemoo Choi, Jaewoong Choi, Myungjoo Kang
TL;DR
This paper unifies OT-based adversarial generative models, analyzes their training dynamics, and introduces a simple refinement method that significantly improves their performance, achieving state-of-the-art FID scores on CIFAR-10 and CelebA-HQ-256.
Contribution
It provides a unified framework for OT-based adversarial models, offers a detailed analysis of their training, and proposes a novel refinement method that enhances generative quality.
Findings
Achieved FID of 2.51 on CIFAR-10
Achieved FID of 5.99 on CelebA-HQ-256
Outperforms previous OT-based adversarial models
Abstract
Optimal Transport (OT) problem aims to find a transport plan that bridges two distributions while minimizing a given cost function. OT theory has been widely utilized in generative modeling. In the beginning, OT distance has been used as a measure for assessing the distance between data and generated distributions. Recently, OT transport map between data and prior distributions has been utilized as a generative model. These OT-based generative models share a similar adversarial training objective. In this paper, we begin by unifying these OT-based adversarial methods within a single framework. Then, we elucidate the role of each component in training dynamics through a comprehensive analysis of this unified framework. Moreover, we suggest a simple but novel method that improves the previously best-performing OT-based model. Intuitively, our approach conducts a gradual refinement of the…
Peer Reviews
Decision·ICLR 2024 poster
- Background knowledge is contained well in the paper. - Existing OT-based methods are analyzed well under the proposed generalizing framework. - An additional scheduling method is proposed for mitigating a drawback of the cost function while boosting the benefit.
- It is not easy to understand without expertise in OT. It would be much easier to read if one or two lines of descriptions comparing each term and notation with those of the regular GAN setup were provided. - The analysis is good, but the benefit of the proposed method (UOTM-SD) is not clearly shown. - Experiments are limited to low-dimensional datasets which is not practical enough. - Some parts are not clear which are described in the Questions below.
By studying the considered adversarial models through a unified framework, this paper provides **interesting comparative insights** on their experimental performance. These insights might be valuable for future research in this area, highlighting the value of OT map models. These insights are **well illustrated** thanks to toy experiments, and the paper is overall **clear and easy to read**. The resulting proposed model refinements provide **improvements in generative performance and robustness
The paper suffers from three main weaknesses that, together, make me believe that it remains under the acceptance threshold. I look forward to discussing with the authors and other reviewers on this topic. ### Significance of the Proposed Framework As it is described in Algorithm 1, the proposed framework is a useful framework for exposition and experimental design, but the significance of this contribution is limited. - It is a straightforward generalization of the already established lookali
The originality of the paper mainly comes from theorems 3.1 and 4.1, which consolidate the recently proposed UOTM method. These theorems pave the way of the proposed method that addresses the tau-sensitivity problem. Clarity of the paper looks good to me overall, but there are some places I'm confused about.
The experiment results look promising as well. It would be great if the proposed method is applied on higher solution image dataset to showcase the image generation quality.
Videos
Taxonomy
TopicsGenerative Adversarial Networks and Image Synthesis · Adversarial Robustness in Machine Learning · Model Reduction and Neural Networks