Lectures by Prof. HT Kung and Prof. Changwen Chen

Title: Speeding Up Inference for Deep Neural Networks Browsing Unstructured Social Media Feeds on Mobile Devices Date and Time:9:00 -11:30 am, June 27, 2016 Venue: Lecture Hall, Office Building, Software Campus

       6月27日上午,来自哈佛大学的孔祥重(HT Kung)教授和来自纽约州立大学布法罗分校的陈长汶(Chang Wen Chen)教授为山东大学计算机科学与技术学院带来了两场精彩的学术报告。陈宝权院长为孔祥重教授颁发了杰出讲者的聘书;报告结束后,两位教授在陈宝权院长的陪同下参观了交叉研究中心。
       孔祥重教授是哈佛大学计算机科学与电机工程学院比尔盖茨的教授。他的研究兴趣有计算机系统、网络、传感器和无线通信,目前则主要专注于机器学习、高性能计算和物联网。他带来的报告Speeding Up Inference for Deep Neural Networks介绍了深度神经网络及其应用。值得一提的是,孔祥重教授也是孔子的嫡孙,与山东颇有渊源,本次访问山东是他回到孔子的故乡,感慨良多,报告中他使用孔子的名言“述而不作,信而好古”来教育同学们对待科研要勇于创新,精益求精。

        陈长汶教授是纽约州立大学布法罗分校计算机科学与工程系教授,长期从事计算机视觉及医学图像分析、图像视频信号编码与通讯、无线通信与网络等领域的教学和研究工作,在这些领域取得了一系列开创性的成果。他的报告Browsing Unstructured Social Media Feeds on Mobile Devices主要介绍了针对社交网络多媒体数据的信息提取、总结与展示的重要技术。




Speeding Up Inference for Deep Neural Networks
We know that deep neural networks can classify image objects and others with high accuracy. However, in feedforward inference, layer-by-layer processing can incur long delays, which would be intolerable for real-time applications such as millimeter wave antenna control for emerging 5G cellular systems.        
We observe that not all data items are equal in their recognition difficulties.  In particular, some samples may be relatively easy, in the sense that a deep neural network can quickly classify them via early exit, thereby skipping all later layers  to allow sped-up inference. This presentation will describe Dynamic Adaptation during Testing (DAT), a method that can exploit this observation, by automatically configuring early-exit criteria. By adapting to the given test set at hand, DAT can  significantly shorten inference time without retraining the network. We have evaluated the DAT method by augmenting a well-studied network (ResNet). We demonstrate that DAT can automatically shorten inference latency for easy test samples by 7.4x and for hard samples by 2.8x, using the same pre-trained network.  DAT is joint work with Harvard graduate students, Brad McDanel and Surat Teerapittayanon.
Browsing Unstructured Social Media Feeds on Mobile Devices
This talk addresses several important technical issues in browsing heterogeneous unstructured multimedia feeds on consumer mobile devices derived from social network contents. We will first present several pressing technical challenges associated with creating a browsing system that can summarize information overloading unstructured social media feeds and produce a novel GIST, namely, Graphical Intelligent Semantic Transform, for effective and visually pleasing browsing on a mobile device by the social media users. We will then illustrate innovative solutions to solving a suite of interdisciplinary problems associated with developing such a system. Preliminary results will be shown to demonstrate the feasibility of creating such a GIST for browsing information overloading social media feeds on consumer mobile devices.
Distinguished Lecture Series (DLS) Program:
Since 2015, the School of Computer Science and Technology, and the School of Software of Shandong University have launched the Distinguished Lecture Series Program that features internationally acclaimed scholars to speak about the frontier in both scientific research and industrial development    in the fast developing computing and software engineering field. The DLS aims to promote academic exchange and raise the visibility of the schools. Each year, no more than ten scholars are honored to speak at the DLS.