学　术

　　人力资源部教师发展中心“学术沙龙”活动特别邀请澳大利亚莫纳什大学Wray Buntine教授、杜岚博士和常晓军博士来校交流，具体安排如下，欢迎广大师生参加。

　　一、时  间：2019年3月27日（周三）上午9:30

　　二、地  点：清水河校区　宾诺咖啡

　　三、主持人：徐增林 教授 计算机科学与工程学院（网络空间安全学院）

　　四、活动安排：

　　报告一

　　主　题：Nonparametric Methods for Classification and Non-negative Matrix  factorisation

　　主讲人：澳大利亚莫纳什大学 Wray Buntine 教授

　　交流内容：

　　Nonparametric methods were the poster child of Machine Learning in the 2000s, and deep neural networks have now taken that role.  In this talk I'll review some of our more successful methods and briefly give results. In classification, we have developed Bayesian network classifiers using non-parametric smoothing that, when ensembled, substantially outperform XGBoost, the acknowledged state-of-the-art leader in applied data science.  For discrete matrix factorisation we have developed non-parametric methods that allow us to deal with bursty data, to incorporate prior information like word embeddings or document meta-data.  For these algorithms we use carefully tuned Gibbs sampling using augmentation and they generally perform at state-of-the art.  In one case where we have tested it, a multi-core implementation (with 8 CPUs) of Gibbs sampling substantially outperformed a stochastic variational inference algorithm for the same task.  As a community we know deep neural networks give us a whole new range of capabilities that should outperform our methods, but I believe it’s important to understand the state of the art beforehand.

　　主讲人简介：

　　Wray Buntine现任莫纳什大学教授，曾是数据科学硕士基金会主管并领导机器学习组。曾工作于NICTA Canberra，赫尔辛基信息技术研究所(负责语义研究项目)，NASA Ames研究中心，加州大学伯克利分校和谷歌。他在文档文本分析的概率方法、社交网络、数据挖掘和机器学习等领域做理论和应用性工作，H指数39，Google引9800多次。他因将贝叶斯方法和贝叶斯网络模型引入机器学习领域及率先使用集成方法而被广泛认可。他还是期刊Data Mining and Knowledge Discovery, Knowledge and Information Systems和Behaviormetrika的编辑，任职于亚洲机器学习会议（ACML）的指导委员会，且是UAI, AAAI, IJCAI, ACML, AISTATS等会议的高级程序委员会成员。

　　报告二

　　主　题：Learning Interpretable Topic Structure from Texts

　　主讲人：莫纳什大学 杜岚博士　　

　　交流内容：

　　One important task of topic modelling for text analysis is to improve the model’s interpretability. We are interested in analysing the fine-grained thematic structure within each individual topic, for example, the inter/intra topic structure and the hierarchical topic structure. In this talk, I will focus on how to leverage either auxiliary information or deep structures for learning words distributions, and present: WEDTM, a model that can discover inter topic structures with topic hierarchies and intra topic structures with sub-topics by leveraging word embeddings, published in ICML 2018; 3) Dirichlet Belief Nets, a new multi-layer generative process on per-topic word distributions, published in NeurIPS 2018.

　　主讲人介绍：

　　杜岚博士目前是莫纳什大学信息技术系的数据科学讲师和数据科学硕士课程总监。他曾工作于麦考瑞大学计算系的语言技术组。他的主要研究兴趣是统计机器学习及其在文本分析上的应用，图表示学习，关系学习和社交网络分析等等。他在包括NIPS,ICML,AISTATS,AAAI,ACL,TPML的顶级会议或期刊上发表过超过35篇文章。他曾在机器学习，数据挖掘和自然语言处理等领域的很多顶级会议（如NIPS, ICML, AISTATS, AAAI, IJCAI, ACL, EMNLP, NAACL, ICRL）的程序委员会任职。　　

　　报告三

　　主　题：面向视频目标切割的强化切割智能体学习

　　主讲人：莫纳什大学 常晓军博士

　　交流内容：

　　Video object segmentation is a fundamental yet challenging task in computer vision community. In this paper, we formulate this problem as a Markov Decision Process, where agents are learned to segment object regions under a deep reinforcement learning framework. Essentially, learning agents for segmentation is nontrivial as segmentation is a nearly continuous decision-making process, where the number of the involved agents (pixels or super pixels) and action steps from the seed (super)pixels to the whole object mask might be incredibly huge. To overcome this difficulty, this paper simplifies the learning of segmentation agents to the learning of a cutting-agent, which only has a limited number of action units and can converge in just a few action steps. The basic assumption is that object segmentation mainly relies on the interaction between object regions and their context. Thus, with an optimal object (box) region and context (box) region, we can obtain the desirable segmentation mask through further inference. Based on this assumption, we establish a novel reinforcement cutting-agent learning framework, where the cutting agent consists of a cutting-policy network and a cutting execution network. The former learns policies for deciding optimal object-context box pair, while the latter executes the cutting function based on the inferred object-context box pair. With the collaborative interaction between the two networks, our method can achieve the outperforming VOS performance on two public benchmarks, which demonstrates the rationality of our assumption as well as the effectiveness of the proposed learning framework.

　　主讲人简介：

　　常晓军博士是澳大利亚莫纳什大学克莱顿校区信息技术学院的讲师。他被评选2019-2021年的ARC Discovery Early Career Researcher Award (DECRA) Fellow。博士毕业于悉尼科技大学人工智能中心和工程与信息技术学院，曾在卡内基梅隆大学计算机科学学院从事学术研究工作，合作导师为Alex Hauptmann教授。他的主要研究兴趣为多种信号（视觉，声学，文本），以便在无约束或监视视频中进行自动内容分析，H指数25，Google引用2100多次。发明的系统在各种国际比赛中（如TRECVID MED，TRECVID SIN和TRECVID AVS）取得了最佳成绩。

　　主办单位：

　　人力资源部教师发展中心

　　承办单位：

　　计算机科学与工程学院（网络空间安全学院）、统计机器智能与学习实验室（SMILE Lab）

　　　　　　　　　　　　　　　　　　　　　人力资源部教师发展中心

　　　　　　　　　　　　　　　　　　　　　　 2019年3月26日