This website requires JavaScript.

Learning Energy-Based Models With Adversarial Training

Xuwang YinShiying LiGustavo K. Rohde
Dec 2020
摘要
We study a new approach to learning energy-based models (EBMs) based onadversarial training (AT). We show that (binary) AT learns a special kind ofenergy function that models the support of the data distribution, and thelearning process is closely related to MCMC-based maximum likelihood learningof EBMs. We further propose improved techniques for generative modeling withAT, and demonstrate that this new approach is capable of generating diverse andrealistic images. Aside from having competitive image generation performance toexplicit EBMs, the studied approach is stable to train, is well-suited forimage translation tasks, and exhibits strong out-of-distribution adversarialrobustness. Our results demonstrate the viability of the AT approach togenerative modeling, suggesting that AT is a competitive alternative approachto learning EBMs.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?
0
被引用
笔记
问答