/

En

1.Theme

Doctoral Forum

2.Schedule

Time:November 01 15:40-17:40
Location:第二会议厅(B+C)
Time Speaker Headline Host
15:40-15:50 王语霖 Dynamic Deep Networks for Reducing Spatial Redundancy 彭玺、苏航、白慧慧
15:50-16:00 代克楠 长时目标跟踪的在线更新
16:00-16:10 汪润中 Learning Lawler's Quadratic Assignment Problem by Neural Graph Matching Network
16:10-16:20 邓志杰 Bayesian Deep Learning: Methods and Applications
16:20-16:30 欧阳熹 基于双采样注意力机制网络的新冠肺炎诊断
16:30-16:40 方杰民 Neural Architecture Search with High Flexibility and Efficiency
16:40-17:30 王亮,山世光,韩军伟,黄高及部分学生代表 会议研讨

3.Proposed lecture content and introduction of invited speakers

Invited Report 1:Dynamic Deep Networks for Reducing Spatial Redundancy
Summary: The performance of deep networks (e.g., ConvNets and Vision Transformers) generally improves when fueled with high resolution inputs. However, this often comes at a high computational cost and high memory footprint. This talk will introduce our recent works on alleviating this inefficiency through the lens of reducing spatial redundancy. First, a reinforcement learning-based approach, GFNet, will be presented, which dynamically identifies and attends to the most informative regions of each individual image. Then, this framework will be extended to processing videos. Our methods can be implemented on top of state-of-the-art light-weighted networks (e.g., MobileNet-V3, EfficientNet, RegNet and TSM), and effectively improve their efficiency, both theoretically and empirically. Finally, this talk will introduce an efficient dynamic Vision Transformer (DVT), where the number of representative tokens is adaptively configured conditioned on the inputs.
王语霖

Speaker:王语霖,清华大学自动化系三年级直博生,导师为吴澄院士和黄高助理教授,研究兴趣为深度学习模型的高效训练和推理方法, 尤其关注动态神经网络的设计和训练。在TPAMI、NeurIPS、ICLR、ICCV、CVPR上发表学术论文10篇,其中第一作者6篇,发表于ICCV、CVPR的两篇论文被 选为“Oral Presentation”,科研工作在GitHub网站上的开源代码已获得800余颗star。 担任TPAMI、IJCV、T-Cybernetics、CVPR、ICCV、NeurIPS、 ICLR等顶级期刊和会议的审稿人,获评CVPR-2021“Outstanding Reviewer”。
个人主页: www.rainforest-wang.cool



Invited Report 2:长时目标跟踪的在线更新
Summary: 相比于短时目标跟踪,长时目标跟踪视频更长,目标消失出境更加频繁,目标外观信息也更加多样,更接近实际应用。大多数排名靠前的长时跟踪算法都采用不更新的策略, 这主要是因为长时目标跟踪的在线更新充满了不确定观测的风险。为了解决这个问题,我们首先提出了一个更新控制器,将目标的判别信息,几何信息,外观信息进行时序封装, 送入长短时记忆网络学习当前帧是否可以更新,极大了提升了更新的准确率。其次提出了一个长时跟踪框架,使短时跟踪器可以轻易的嵌入变成长时跟踪器。该方法让我们 在多个长时跟踪数据集上都取得了非常领先的结果。
代克楠

Speaker:代克楠,大连理工大学三年级硕士,师从卢湖川教授。研究方向目标跟踪,目标检测,分割等计算机视觉方向。硕士至今发表两篇一作CVPR, 一篇一作ICCV,一篇二作ACM MM,其中两篇CVPR均为Oral,其中一篇还获得了CVPR2020最佳论文提名,Google scholar引用共390余次。多次参加目标跟踪国际竞赛(VOT), 并两次获得长时组冠军。



Invited Report 3:Learning Lawler's Quadratic Assignment Problem by Neural Graph Matching Network
Summary: Graph matching involves combinatorial optimization based on edge-to-edge affinity matrix, which can be generally formulated as Lawler’s Quadratic Assignment Problem (QAP). This paper presents a QAP network directly learning with the affinity matrix (equivalently the association graph) whereby the matching problem is translated into a vertex classification task. The association graph is learned by an embedding network for vertex classification, followed by Sinkhorn normalization and a cross-entropy loss for end-to-end learning. We further improve the embedding model on association graph by introducing Sinkhorn based matching-aware constraint, as well as dummy nodes to deal with unequal sizes of graphs. To our best knowledge, this is the first network to directly learn with the general Lawler’s QAP. In contrast, recent deep matching methods focus on the learning of node and edge features in two graphs respectively. We also show how to extend our network to hypergraph matching, and matching of multiple graphs. Experimental results on both synthetic graphs and real-world images show its effectiveness. For pure QAP tasks on synthetic data and QAPLIB benchmark, our method can perform competitively and even surpass state-of-the-art graph matching and QAP solvers with notable less time cost.
汪润中

Speaker:汪润中,上海交通大学三年级博士生,师从严骏驰副教授和杨小康教授,主要研究方向是机器学习、组合优化以及相关技术在计算机视觉中的应用, 尤其是深度学习图匹配技术在计算机视觉任务中的应用。汪润中已在CVPR、ICCV、NeurIPS、TPAMI上发表第一作者论文7篇,担任CVPR、ICCV、NeurIPS、ICLR、ECCV、 AAAI等会议的审稿人。



Invited Report 4:Bayesian Deep Learning: Methods and Applications
Summary: Despite its immense popularity, deep learning is known to face limitations in various aspects, e.g., (i) deep learning models suffer from overconfidence and lack a proper mechanism to quantify the uncertainty, which is ubiquitous and essential in the physical world; (ii) deep learning models are purely data-driven, unable to effectively benefit from prior knowledge. In this talk, I will introduce how to address these issues in the light of Bayesian deep learning. I start by the introduction of the challenges in the Bayesian treatment of DNNs and then provide strategies to tackle them. I will cover our recent works on Bayesian fine-tuning, lightweight Bayesian modeling, architectural uncertainty, as well as some advanced topics like the connections between kernels and deep models. I will also demonstrate the benefits of Bayesian deep learning by some application examples.
ZhiJie Deng

Speaker:I am ZhiJie Deng,a fifth-year Ph.D. student of TSAIL Group in the Department of Computer Science and Technology, Tsinghua University, advised by Prof. Bo Zhang and Prof. Jun Zhu. I was a visiting student from July 2016 to September 2016 in the Machine Learning Department, Carnegie Mellon University, advised by Prof. Eric P. Xing. My research interest includes deep learning and statistical methodology. I have published 8 papers on top-tier ML/CV conferences like NeurIPS, CVPR, ICCV, and ATC.
My homepage is at https://ml.cs.tsinghua.edu.cn/~zhijie/.



Invited Report 5:基于双采样注意力机制网络的新冠肺炎诊断
Summary: 新型冠状病毒(COVID-19)仍在全球范围内传播,已经感染了超过两亿人。核酸检测仍然是新冠肺炎诊断的金标准,与此同时,胸部计算机断层扫描(Computed Tomography, CT)也可以实现对 疑似新冠患者的快速筛查。为此,我们开发了一个基于双重采样注意力机制的深度神经网络,可以通过CT图像自动诊断区分社区获得性肺炎和新冠肺炎。具体而言,我们提出了一种采用了新颖的注意力机制 的三维卷积网络,使得网络在做出诊断决定时像放射科医生一样把关注点放在肺部的感染区域。同时我们发现部分新冠肺炎和社区获得性肺炎之间存在感染面积大小的不平衡。因此我们提出了一个双采样策略 来减轻这种不平衡对网络学习的影响。我们的方法在来自8家医院的大规模新冠CT数据中进行了验证评估。
欧阳熹

Speaker:欧阳熹上海交通大学博士四年级学生,导师为王乾研究员和上海科技大学沈定刚教授。研究方向为医学图像分析,主要研究深度学习和弱监督学习在医疗图像分析中的应用。 先后在TMI、MICCAI等医学图像领域顶刊和顶会发表论文5篇,论文在Google Scholar引用560次。



Invited Report 6:Neural Architecture Search with High Flexibility and Efficiency
Summary: The emergence of deep neural networks (DNNs) has dramatically boosted the development of various fields in Artificial Intelligence, especially computer vision. As DNNs have more and more applications in real-life scenarios, the efficiency of DNNs should be taken into considerations to satisfy the hardware properties. As one of the most important techniques to promote DNN efficiency, neural architecture search (NAS) has been widely studied and applied in recent years. This talk will mainly introduce our research works about two key topics in neural architecture search, i.e., flexibility and efficiency. In our works, the flexibility is enhanced by introducing densely connected search spaces but with low search cost; the efficiency is promoted, especially in downstream tasks, via the proposed parameter remapping mechanism.
Jiemin Fang

Speaker:Jiemin Fang is a PhD candidate at the Institute of Artificial Intelligence and School of Electronic Information and Communications, Huazhong University of Science and Technology. He is supervised by Prof. Wenyu Liu and Prof. Xinggang Wang. He received the B.E. degree from EIC, Huazhong University of Science and Technology in 2018. His research interests include AutoML and efficient deep learning. He has published research papers on journals and conferences, including TPAMI, CVPR, ICLR, ICCV, NeurIPS, etc. He has served as a research intern in Huawei Cloud and Horizon Robotics.
Visit https://jaminfong.cn/ for more information.





4.The organizer

彭玺

个人简介:彭玺(pengxi@scu.edu.cn)四川大学计算机学院教授,博导,国家特聘青年教授;于2013年12月毕业于四川大学计算机学院;2014年至2017年,就职于新加坡科技局资讯通信研究院 (Institute for Infocomm, A*STAR)担任研究员和联合项目主管(CO-PI);2017年被四川大学引进任特聘研究员;2018年入选四川省特聘教授;2019年入选"国家特聘青年教授", 同年9月破格转聘为教授。彭玺的主要研究方向包括表示学习基础理论及其在多媒体计算、视觉计算、自然语言处理、图像处理等领域中的应用。他共计在中国计算机学会A类推荐期刊/会议 (ICML,NeurIPS,CVPR,ICCV,AAAI,IJCAI、TPAMI、IJCV、TIP、TIFS等)上发表学术论文60余篇。发表的一作/通信论文中包含多篇ESI热点论文(前1‰)和高被引论文(前1%); 获教育部霍英东青年教师奖(二等,全国15人)等多个奖励;指导学生获全国互联网+等奖励多项;是第十届中国科协全国代表大会代表。



苏航

个人简介:苏航(suhangss@mail.tsinghua.edu.cn),清华大学计算机系副研究员,主要研究对抗机器学习和鲁棒视觉计算等相关领域,发表CCF推荐A类会议和期刊论文30余篇, 谷歌学术论文引用2800次,获得ICME铂金最佳论文、MICCAI青年学者奖和AVSS最佳论文等多个学术奖项,曾率队在NeurIPS2017对抗攻防等多个国际学术比赛中获得冠军。现任中国图像图形学会青工委执委、 VALSE执行AC委员会主席,担任NeurIPS21的领域主席(Area Chair)、AAAI22 Workshop Co-Chair,并在多次ICML等顶级国际会议上作为分论坛主席组织对抗学习专题研讨。





白慧慧

个人简介:白慧慧(hhbai@bjtu.edu.cn)北京交通大学教授、博士生导师、信息科学研究所副所长。2001年及2008年获得北京交通大学电子与信息技术学士学位及信号与信息处理专业博士学位。 2008年起在北京交通大学计算机与信息技术学院工作。2014年获国家公派留学资助赴加拿大西蒙菲莎大学访问。主要研究方向是图像视频编码和增强等。已发表学术论文60余篇,包括IEEE汇刊与计算机学会推荐会议论文等 20余篇。获美国专利授权1项、澳大利亚创新专利授权1项、国家发明专利授权10项。主持了国家自然科学基金项目、北京市自然基金项目、江苏省自然科学基金项目等。获北京市科学技术奖一等奖、中国产学研合作创新成果奖 二等奖、山西省科学技术奖三等奖、山西省高校科学技术一等奖、北京图象图形学学会优秀导师奖。入选北京高等学校“青年英才计划”、微软亚洲研究院“铸星计划”、CCF-腾讯犀牛鸟创意基金、APSIPA Distinguished Lecturers。