Mingxuan's paper introduces a fresh perspective on generative models, which are used in popular machine learning applications like ChatGPT and Stable Diffusion. Generative Adversarial Net (GAN) is a commonly used technique for training generative models, and has traditionally been viewed as a process of minimizing divergence. However, in this paper, we suggest that it is more beneficial to think of GAN training as an Ordinary Differential Equation (ODE) process. This new approach allows for the development of better generative loss functions, leading to more effective GAN training. Our theoretical findings are supported by experimental results. Mingxuan's paper is a collaboration with Prof. Zhanxin Zhu from Peking University.
Daniel's paper proposes a new way to estimate the density of truncated datasets, which are commonly encountered in machine learning tasks. For example, if we only have data on crime occurrences within a certain city boundary, we have a truncated dataset. Previous methods for dealing with truncated datasets required explicit knowledge of the boundary function, but the method proposed in this paper only requires approximate knowledge of the boundary, based on samples on the boundary. Ours approach outperforms existing methods on both simulated and real-world datasets, even when the boundary is not well-defined.
(ICML2023), Williams, D. J., Liu, S., Approximate Stein Classes for Truncated Density Estimation, ICML2023, to appear. (to be uploaded to Arxiv soon)
ICML 2023 will be held in Honolulu, Hawai'i from July 23rd - July 29th.