2020年9月17日星期四

CityU SDSC Seminar - Big Data, Deep Learning, and Federated Learning Research at Baidu

The CityU Online Seminar named “Big Data, Deep Learning, and Federated Learning Research at Baidu” was organized by The School of Data Science (SDSC) and The Hong Kong Institute for Data Science (HKIDS) and was held on 17th Sep 2020.  This talk invited Prof. Dejing Dou (Head of Big Data Lab (BDL) and Business Intelligence Lab (BIL) at Baidu Research) to share “Open and Inclusive” and Baidu AutoDL which used deep learning to design and train deep learning models, including Neural Architecture Search, transfer learning, and interpretability of deep learning.



Prof. Dou said Big Data and AI would become the “engines” of our times! Then he briefed Big Data’s 3V that is Volume, Velocity (Speed) and Variety (Complexity).  



And then Prof. Dou introduced Big Data Lab at Baidu Research which was founded in collaboration with UN in 2014. Their research areas included Neural Architecture Search (NAS), Transfer Learning, Interpretability of DNN and Federated Learning.  He also said AI applications would be industrialized, standardized, modularized and automated.  



Because of current limitations in Deep Learning (DL), Prof. Dou expected the future DL would be follows:

      -          Scalable, automated and customized design
-          Ubiquitously deployed on low-cost and heterogeneous computing hardware ends
-          Self-organizing, self-adaptive, self-evolving; reduced dependence on data
-          Enhanced interpretability; robust

And then he briefed three key techniques in AutoDL and they were Self-design Architectures, Self-adaptive to Tasks and Self-adjusting to Environments.


After that Prof. Dou discussed Auto Deep Learning and to be global trend such as Google (Cloud AutoML & Differentiable Neural Architecture Search (DARTS)), Microsoft (MS Customer Vision, MS Azure ML), Amazon (Amazon ML) and Salesforce (Transmogrif AI). Baidu AutoDL were introduced by Prof. Dou with three components including AutoDL Design, AutoDL Transfer and AutoDL Edge.

Baidu AutoDL included:

1.      Continuing the research excellence

2.      Cross-modal Transfer Learning

3.      Continuous Integration with PaddleHub Finetune API

4.      Continuous Integration with EasyDL



Prof. Duo also mentioned Interpretable DL including Explain Models, Explain Training Procedures and Integration with PaddleX.  He then briefed the milestones of Baidu Auto DL from 1.0 to 3.0.



Finally, Prof. Duo introduced research in Baidu Business Intelligence Lab (BIL) and its advance mobile AI technologies including Urban Quantitative Analysis, Urban Cognitive Computing, Intelligent Online Map Service and Emerging AI Methods & Applications. 



Lastly, he summarized that Big Data and AI are the core power in this era. Big Data Lab aimed to transform big data into practice knowledge.  He concluded “Open and Inclusive AI” could be applied in different areas.



Q&A session performed at the end.

Prof.  asked interpretability of DL model into physic meaning. Prof. Duo said most people focused on prediction results but explanation of that results (input and output) were also important.

Then I asked how to verify DL model validity in Small Data. Prof. Duo said to use Transfer Learning from Big Data model to fit your small data. If it is simple, just use regression rather than DL. 



Reference:

SDSC - https://www.sdsc.cityu.edu.hk/

HKIDS - https://www.cityu.edu.hk/hkids/

Poster - https://www.cityu.edu.hk/sdsc_web/seminar/2020-09-17_sdsc_seminar.pdf

Other CityU’s AI seminar:

20200819: CityU Staff Seminar - Copyright As An Obstacle to AI Development - https://qualityalchemist.blogspot.com/2020/08/cityu-staff-seminar-copyright-as.html

20191101: CityU & PolyU Jointly Seminar on Attacks and Defenses on Machine Learning Services - https://qualityalchemist.blogspot.com/2019/11/cityu-polyu-jointly-seminar-on-attacks.html

20190912: CityU Seminar on the challenges and responses of AI to Public Management - https://qualityalchemist.blogspot.com/2019/09/cityu-seminar-on-challenges-and.html

20190814: CityU & NVIDIA Technology Sharing - AI Application and Research - https://qualityalchemist.blogspot.com/2019/08/cityu-nvidia-technology-sharing-ai.html

20190409: CityU & TFI Seminar – AI for Better Health and Wealth - https://qualityalchemist.blogspot.com/2019/04/cityu-tfi-seminar-ai-for-better-health.html

20190409: AI and E-Commerce: Ethical Conflicts《人工智能與電子商貿:倫理對弈》 - https://qualityalchemist.blogspot.com/2019/04/ai-and-e-commerce-ethical-conflicts.html

20190322: CityU President's Lecture series - Big Data Analysis & AI: Opportunities & Challenges - https://qualityalchemist.blogspot.com/2019/03/cityu-presidents-lecture-series-big.html

20190227: CityU Seminar on Design of Accident Prevention System for LWR using ANN and HS Simulator - https://qualityalchemist.blogspot.com/2019/02/cityu-seminar-on-design-of-accident.html

20190128: CityU Distinguished Lecture on AI Enabled Personalized Theranotics - https://qualityalchemist.blogspot.com/2019/01/cityu-distinguished-lecture-on-ai.html

20190122: CityU Seminar on The First Step for AI-based Human-Like Language Understanding – Sentiment Analysis of Text - https://qualityalchemist.blogspot.com/2019/01/cityu-seminar-on-first-step-for-ai.html

20181227: CityU Distinguished Lecture - How to Make an Artificial Vision System Smart? - https://qualityalchemist.blogspot.com/2018/12/cityu-distinguished-lecture-how-to-make.html

20181114: CityU the 1st Workshop on Financial Data Analysis - https://qualityalchemist.blogspot.com/2018/11/cityu-1st-workshop-on-financial-data.html

20180804: HKSQ AGM Quality Innovation Seminar on “Extenics and AI Application” - https://qualityalchemist.blogspot.com/2018/08/hksq-agm-quality-innovation-seminar-on.html

20180607: CityU Workshop on AI in the Era of Big Data - https://qualityalchemist.blogspot.com/2018/06/cityu-workshop-on-ai-in-era-of-big-data.html


沒有留言:

LinkWithin

Related Posts with Thumbnails