Improved Generalization of cGAN using Vicinal Estimation and Early Stopping

Ki Joung Jang, Ganguk Hwang

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Supervised Learning, Conditional density estimation, Generative Adversarial Network, Vicinal Risk Minimization

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose a new conditional Generative Adversarial Network (cGAN) model and obtain a generalization error bound of the proposed model which is independent of the output dimensionality.

Abstract: The problem of generating high-dimensional distributions has been known as a difficult problem in machine learning due to the Curse of Dimensionality: the higher the dimensionality is the more the empirical data deviates from its original distribution even for a large number of samples. Along with the Curse of Dimensionality, the generalization of conditional density estimation (CDE) suffers from so-called Lack of Conditional Samples: the number of data for each conditional density is usually much smaller than the number of samples or no data is avaiable for some conditional densities. To overcome these difficulties, we introduce the concept of Vicinal Estimation (VE) which is shown to be useful in estimating conditional densities. With VE we propose a conditional Generative Adversarial Network (cGAN) model and analyze theoretically that the generalization error of our model is independent of the dimensionality of the output. We also show that our theoretical analysis holds in practice through experiments.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3273