June 18-21, 2019
Beijing, CN


此演讲使用中文 (This will be presented in Chinese)

陈玉荣 (Intel)
13:1013:50 Friday, June 21, 2019
模型与方法 (Models and Methods)
Location: 多功能厅6A+B (Function Room 6A+B)

必要预备知识 (Prerequisite Knowledge)

  • A general knowledge of deep learning, convolutional neural network (CNN) architecture, deep neural network (DNN) quantization, and optimization

您将学到什么 (What you'll learn)

  • Learn three key ways to enable deep learning inference at the edge devices from DNN algorithm design perspective

描述 (Description)



我们的高效CNN算法设计主要集中在物体检测任务上,因为它是视觉识别理解的最基本任务。我们设计了几种新颖的CNN网络结构来实现高效准确的物体检测,包括最先采用卷积层特征聚合技术来提高小目标物体检测准确度的HyperNet (CVPR’16),最早提出特征金字塔表示特征进行物体检测工作之一的RON (CVPR’17),最先提出从头开始训练的物体检测器DSOD (ICCV’17),以及在DSOD基础上提出的轻量级物体检测器Tiny-DSOD (BMVC’18)。其中,DSOD及Tiny-DSOD已被使用于英特尔Unite和Altia PanaCast Systems的智能视频会议系统。

除了具有针对性的高效网络结构设计外,DNN模型压缩是在边缘端实现深度学习推断的另一项关键技术。我们研究开发了一整套业界领先的不损失模型推断准确度的深度神经网络压缩技术,包括DNS (NIPS’16),Network Slimming (ICCV’17)等多种DNN网络剪枝技术,以及INQ (ICRL’17),Network Sketching (CVPR’17),MLQ (AAAI’18),ELQ (CVPR’18)等低比特DNN模型量化技术。尤其是我们最近提出的ELQ超低比特量化技术在ImageNet数据集上实现了三值和二值CNN模型量化的领先准确度。我们的这套技术已经获得十余项美国专利申请批准,并作为英特尔研究院的一项重大技术转化,帮助英特尔产品部门来提升英特尔平台的深度学习推断软硬件加速能力。

DNN模型压缩技术通常要进行重训练,来确保模型推断准确度(尽量)不损失。重训练是比较耗时的,并且在很多时候我们不能获得用户的数据进行重训练。为了解决这个问题,我们还提出了一套部署时DNN网络结构优化技术,包括网络解耦Network Decoupling (BMVC’18), 复合二值分解网络CBDNet (AAAI’19),少量知识蒸馏等,它的优点是不需要重训练及深度学习专家参与,就可以在部署DNN时,对其进行加速。

The Cognitive Computing Lab at Intel Labs China focuses on cutting-edge technology research in visual cognition (including perception, recognition, understanding, and cognition) and machine learning, by which it enables innovative applications and user experience for Intel Intelligent Computing. In the past three years, the company has focused on the research of deep learning-based visual understanding and synthesis, boldly carrying out technical innovations, and also drove technology translation, achieving many research outcomes. In the domain of computer vision, machine learning and deep learning, the company holds 60+ issued/pending patents and has published nearly 30 papers at top CV/ML/AI conferences. Meanwhile, it also completed 13 technology translations to Intel’s product divisions, including three major ones. These technology translations have impacted Intel’s hardware and software platform design, improving user experience of Intel products. In addition, it has achieved great results in a number of world-class visual recognition and understanding challenges and provided key AI technologies and related support for several Intel brands’ promotion activities, including Yuchun Li artificial intelligence MV, Great Wall natural heritage protection program, and protection of wildlife Siberian tigers. Deep learning has made great breakthroughs in many areas, especially in visual recognition and understanding. It is, however, facing some challenges in terms of training and deployment.

Yurong Chen outlines the company’s state-of-the-art research, such as solving deep learning deployment challenges through algorithmic innovation and deep learning on the edge. These algorithmic innovations include efficient CNN algorithm designs, cutting-edge DNN model compression technologies, and DNN network architecture optimization techniques for deployment stage.

The efficient CNN algorithmic design focuses on the object detection task because it is the most basic task in the visual understanding domain. The company has designed several novel CNN network architectures to achieve efficient and accurate object detection, including HyperNet (CVPR’16), which first used convolutional layer feature aggregation technology to improve detection accuracy of small target objects; RON (CVPR’17), one of the first works proposed to use feature pyramid representation for object detection; DSOD (ICCV’17), the first proposed method to train a model from the beginning for object detection; and the lightweight object detector Tiny-DSOD (BMVC’ 18), which is based on DSOD. Among them, DSOD and Tiny-DSOD have been used in Intel’s intelligent video conferencing systems, the Unite and Altia PanaCast Systems.

Besides efficient network architecture for specific applications, model compression is another key technology for deep learning inference that can run on edge devices. The company has developed a set of industry-leading deep neural network compression techniques that do not impact model inference accuracy, including DNS (NIPS’16), Network Slimming (ICCV’17), and other DNN network pruning techniques, as well as low-bit DNN model quantization techniques, such as INQ (ICRL’17), Network Sketching (CVPR’17), MLQ (AAAI’18), and ELQ (CVPR’18). In particular, recently proposed ELQ ultra-low bit quantization technology achieves the best accuracy of ternary and binary CNN model quantization on the ImageNet dataset. This set of techniques has been issued with more than a dozen US patents and serves as a major technology translation at the Intel Labs, which helps our product divisions to accelerate deep learning inference capacities for hardware and software at the Intel’s platforms.

DNN model compression is usually required to retrain a model so that its inference accuracy would not decrease if possible. Retraining is time consuming. And in many cases, the company just cannot obtain the same user’s data for retraining. In order to solve this problem, it also proposed a set of DNN network architecture optimization techniques in deployment stage, including network decoupling (BMVC’18), composite binary decomposition network CBDNet (AAAI’19), a small amount of knowledge distillation, etc. In terms of advantages, these techniques can accelerate DNN models during deployment, no need of retraining or involving deep learning experts.

Photo of 陈玉荣



Yurong Chen is a principle research scientist and senior research director at Intel and the director of the Cognitive Computing Lab at Intel Labs China, where he’s responsible for driving cutting-edge visual cognition and machine learning research for Intel smart computing. He’s also the co-owner of Intel Labs visual understanding and synthesis program, driving research innovation in smart visual data processing technologies on Intel platforms across Intel Labs. He led the research and development of deep learning-based visual understanding (VU) and leading face analysis technologies to impact Intel architectures/platforms and delivered core technologies to help differentiate Intel products including Intel RealSense SDK, CV SDK, IOT video E2E analytics solutions, and client apps. He led the team to win the Intel China Award (top team award of Intel China) 2016, Intel Labs Academic Awards (top award of Intel labs), and the Gordy Award 2016, 2015, and 2014 for outstanding research achievements on DL-based VU, multimodal emotion recognition and advanced visual analytics. He’s published over 50 technical papers and holds 10+ issued/pending US/PCT patents and 30+ patent applications. He holds a PhD from Tsinghua University, which he followed with postdoctoral research at the Institute of Software, CAS.

Yurong Chen博士是英特尔公司的首席研究科学家和高级研究总监,以及英特尔中国研究院认知计算实验室主任。目前,他负责推进英特尔智能计算的前沿视觉认知(视觉分析和理解)和机器学习研究。他还是英特尔研究院“视觉理解与合成”项目的共同负责人,主导和推动基于英特尔平台的智能视觉数据处理的技术创新。他领导和推动了基于深度学习的视觉理解以及领先人脸分析技术的研究和开发,以此影响英特尔架构/平台设计,并为英特尔实感技术,计算机视觉软件开发包,移动终端应用和物联网端对端视频分析解决方案提供关键技术(人脸检测识别,物体检测,表情识别,深度模型压缩等)。由于在先进视觉分析、多模态情感识别、及基于深度学习的视觉理解方面取得卓越研究成就,他的团队获得了2016年英特尔中国区最高团队奖——英特尔中国荣誉奖,并连续获得了英特尔研究院2014、2015和2016年度全球最高学术奖-戈登•摩尔奖。Chen博士在中国科学院软件研究所完成博士后研究后,于2004年加入英特尔。他于2002年获得清华大学博士学位。他至今已发表顶级学术论文50余篇,拥有10余项美国/国际专利及30多项专利申请。

Comments on this page are now closed.


Lu Jing | 工程师
2019-06-24 20:39 CST