智能感知与计算系列讲座
Lecture Series in Intelligent Perception and Computing 

题    目 （TITLE）：Energy-based adversarial generative models and deep discriminative models

讲 座 人 （SPEAKER）: 屠卓文 教授，加州大学圣地亚哥分校

主 持 人 (CHAIR)：张兆翔 研究员

时    间 (TIME)：2022年10月14日（周五），10：00

地    点 (VENUE)： 腾讯会议ID: 591 900 180 Password: 2022

报告摘要（ABSTRACT）：

In this talk, we will present our line of work by building energy-based adversarial generative models, starting from the 2007 CVPR paper "Learning Generative Models via Discriminative Approaches" where energy-based generative models are learned via adversarial training. We will then present later work by introducing introspective neural networks (INN) where an integrated framework for unsupervised and supervised learning is developed. We aim to build a single model that is simultaneously generative and discriminative. This is achieved by turning standard convolutional neural networks (CNN) into a generic generator using stochastic gradient descent (SGD) Langevin sampling through backpropagation. When followed by iterative discriminative learning, desirable properties of modern discriminative classifiers are directly inherited by the generator. The learned generative model/discriminative classifier is capable of introspection: being able to self-evaluate the difference between its generated samples and the given training data. Our CVPR 2018 work further enhances the introspective neural networks (INN) algorithms by adopting a Wasserstein distance. On the discriminative classification side, INN achieves the state-of-art-class classification and shows its particular robustness against adversarial samples; on the generative modeling side, appealing images are generated for a series of tasks including texture modeling, object modeling, and generation. We will also present our work in the early development of deep convolutional neural networks including deeply-supervised nets (DSN), holistically-nested edge detection (HED), and ResNeXt. In addition, we will briefly go over our recent line of work in building Transformer-based approaches for image classification, object detection, and geometric structure extraction.

报告人简介（BIOGRAPHY）：

Zhuowen Tu is a full professor of Cognitive Science (affiliated with the Department of Computer Science and Engineering), University of California San Diego. Before joining UCSD in 2013 as an assistant professor, he was a faculty member at UCLA. Between 2011 and 2013, he took a leave to work at Microsoft Research Asia. He received his Ph.D. from the Ohio State University and his M.E. from Tsinghua University. He is a recipient of the David Marr Prize award 2003 and a recipient of the David Marr Prize Honorable Mention award 2015. He is a Fellow of the IEEE.