Quality Alchemist (品質煉金術師): CAAI Chinese Congress on Artificial Intelligence 2020 (中国人工智能大会)

2020年8月30日星期日

CAAI Chinese Congress on Artificial Intelligence 2020 (中国人工智能大会) - Day 2

The 6th Chinese Congress on Artificial Intelligence (CCAI2020) which was led by Chinese Association for Artificial Intelligence (CAAI), was held on Aug 29-30, 2020, in Nanjing, China. Since COVID-19 affected globally, the Congress was run both online and offline. The frame of congress named “Intellectual Know for Everything (智周万物)”. I was committee member of the CAAI Extenics Professional Committee (中國人工智能學會可拓學專業委員會) attended the congress online in this time.

Day 2 (30 Aug 2020):

In the beginning, Prof. Yang Qiang (杨强 - CAAI名誉副理事长，微众银行首席AI官) chaired the Morning session for keynotes.

The first speaker was Prof. Yolanda Gil (人工智能促进会现任主席，南加州大学信息科学研究所科研副主任) and her topic was “Thoughtful Artificial Intelligence: Forging a New Partnership for Data Science and Scientific Discovery”. Firstly, she briefed AI application had separated into Data Thread and Knowledge Thread.

Prof. Gil discussed four main points in this talk including Knowledge Technology, AI and its interdisciplinary frontiers and Thought AI.

Then she mentioned the Thoughtful AI using the photo of Freestyle Chess Champion Anson Williams and quoted Gary Kasparov (2010) statement that “Weak human + machine + better process was superior to a strong computer alone and, more remarkably, superior to a strong human + machine + inferior process.” And then she explained seven Thoughtful AI principles.

i) Rationality Principle – From knowledge to behavior

ii) Context Principle – Capture the purpose and significance of tasks

iii) Initiative Principle – Self-driven learning from the scientific record or through independent inquiry

iv) Network Principle – Seek resources from a knowledge web of the scientific record

v) Articulation Principle – Communicating findings to different audiences and respond but also ask and discuss.

vi) Ethical Principle – Behavior that conveys awareness of scope and limitations

vii) Systems Principle – Compositionality, abstraction, connectivity

At the end, Prof. Yolanda Gil gave conclusion on Thoughtful AI in the following four points:

i) Knowledge technologies are increasing important.

ii) AI offers systematic, correct, unbiased approaches and rigorous reporting.

iii) AI will excel at assembling fragmented knowledge about complex systems and pursue interdisciplinary frontiers.

iv) Thoughtful AI will exploit knowledge technologies for effective human-AI partnerships.

The second speaker was Dr. Ming Zhou (周明 - 微软亚洲研究院副院长) and his presentation was “Pre-trained Models in Multi-Lingual and Multi-Modality Tasks” (预训练模型在多语言、多模态任务的进展).

Firstly, Dr. Zhou explained the pre-trained model concept and its application such as for text, auto-encoding, classification, question-answering … etc. Then he explained why pre-trained models need as follows:

i) Pre-trained models embed task-agnostic general knowledge.

ii) Pre-trained models transfer learnt knowledge to downstream tasks.

iii) Pre-trained models supports almost all NLP tasks with state-of-the-art results.

iv) Provide a scalable solution to various applications.

After that he briefed pre-trained models in multi-lingual tasks and multi-modality tasks. And they got a good evaluation results.

Finally, Dr. Zhou summarized their future works included pre-trained models in NN-NLP and explore new tasks, model architectures and empowered with interpretation mechanisms to NN-NLP.

The third speaker was Prof. Tian Qi (田奇 - 华为云人工智能领域首席科学家，教育部长江学者讲座教授) and his topic named “General Vision: Exploration. Practice. Precipitation” (通用视觉：探索.实践.沉淀). He briefed the three key area of AI technology innovation in Huawei Cloud.

After explained different technology under three key area, Prof. Tian mentioned industry application such as mobile quality through AI vision inspection model.

Finally, Prof. Tian introduced Huawei Cloud AI talent training plan included seeking, deposition and practice. They cooperated with different local universities and had different programs.

The fourth speaker was Prof. Chen Xiaoping (陈小平 - 中国科学技术大学，机器人实验室主任、教授) and his topic named “70 years of Turing Hypothesis: Two Types of AI and Closed Challenge” (图灵假说70年：两类AI与封闭性挑战).

Prof. Chen briefed the Turing Hypothesis and explained two types of AI and they were Robot AI and Information Technology (IT) AI. Turing was only focused on IT AI in 1950.

Then Prof. Chen explained the difficulties on modeled grounding because of open scenarios. He added AlphaGo worked in closed scenarios that focused on 362 variables of move.

After that he foresighted the future works of Open Knowledge (OK) as follows:

- Causal reasoning involving grounding to various real-world scenarios

- Combining heterogeneous knowledge systematically on the OK framework

- Systematical techniques of the principles of Disparity (融差性).

Finally, Prof. Chen expected AI development in 2020 to 2025 would be mass application in industry, fast change on foundation research and focused on AI ethics.

The afternoon session I selected to attend “Intelligent Perception and Interaction Forum” (智能感知与交互专题论坛)

The first speaker in the Forum was Prof. Shi Yuanchun (史元春–清华大学计算机系教授，“长江学者”特聘教授) and his presentation topic was “Encoding-decoding Optimization Method for Natural Action Interaction” (自然动作交互的编-解码优化方法). In the beginning, she briefed the human-machine interaction history from CUI (Conversational User Interface) to GUI (Graphical User Interface) and VUI (Voice User Interface) and so on.

And then Prof. Shi explained the Natural Action Interaction’s advantages (such as No instrument, Eyes-free & Muscle memory) and disadvantages (such as Lack of sensing techiques, Recognition accuracy & no. of gestures remembered).

Finally, Prof. Shi proposed Human-Machine Interaction approach through OS, Computer System and User Interface to achieve different applications.

The second speaker was Prof. Wang Yunhong (王蕴红–北京航空航天大学计算机学院教授) and his presentation entitled “Video-based Group Behavior Analysis” (基于视频的群体行为分析). Firstly, she briefed group behaviors in sport videos.

Then she mentioned some key problems including tracing athlete, identify behavior and strategy analysis, etc. After that Prof. Wang briefed the AI algorithms to achieve this purpose.

Finally, Prof. Wang asked three open questions and they were application of interaction, increasing classification power and enhancing the system’s intelligence.

The third speaker was Prof. Wang Liang (王亮–中科院自动化所研究员) and his topic named “Natural Language-driven Pedestrian Search” (自然语言驱动的行人搜索). Prof. Wang firstly explained propose of this research to assist policy to search Suspect through witness description.

And then Prof. Wang explained different AI approaches to handle different description and features to increase the accuracy.

Lastly, Prof. Wang compared with traditional searching that image and language description could be complementary each other.

The fourth speaker was Prof. Liu Guangcan (刘光灿–南京信息工程大学自动化学院教授，国家优青) and his presentation entitled “Universal Low-Rank Learning: Interpretable Deep Learning?” (普适性低秩学习：可解释深度学习？). Prof. Liu reviewed statistical learning which early used in machine learning. Low-rank learning was one of dimensionality reduction methods.

After that Prof. Liu explained Maximum-and-Concatenation Networks (MCN) design for prediction use. Finally, he concluded that low-rank learning should be explore and learn underlying data, as well as, universal low-rank learning’s learn proper transformation (interpretable deep learning) for future data forecasting.

The fifth speaker was Prof. Shan Shiguang (山世光–中科院计算所研究员、博士生导师，国家优青) and his topic was “Human-centered visual perception progress: from recognition to understanding” (以人为中心的视觉感知进展：从识别到理解). Beginning, Prof. Shan explained to observe face for your condition.

And then Prof. Shan extended the understanding of visual technology from physical feathers to psychological feather and then to mental status.

Finally, Prof. Shan stated the development trend of visual perception including task transfer and methodology transfer learning.

The last session was final forum named "Opportunities and Challenges of China Artificial Intelligence Development in the New Situation" (新形势下我国人工智能发展的机遇与挑战) and chaired by Prof. Liu Chenglin (刘成林 - CAAI副理事长，中科院自动化研究所副所长).

Panelists:

Prof. Li Deyi (李德毅 - CAAI名誉理事长，中国工程院院士，CAAI Fellow)

Prof. Zhou Zhihua (周志华 - 南京大学计算机系主任、人工智能学院院长、教授，欧洲科学院外籍院士，ACM/AAAI/IEEE/AAAS/IAPR/CAAI Fellow)

Prof. Yang Qiang (杨强 - CAAI名誉副理事长，微众银行首席AI官，AAAI/ACM/IEEE/CAAI Fellow)

Prof. Zhou Ming (周明 - 微软亚洲研究院副院长)

Prof. Tian Qi (田奇 - 华为云人工智能领域首席科学家，教育部长江学者讲座教授，IEEE Fellow)

Prof. Chen Xiaoping (陈小平 - 中国科学技术大学，机器人实验室主任、教授)

Prof. Liu Chenglin asked panelists opinions on new situation such as US-China trade war and COVID-19.

Prof. Zhou Zhihua said we should have sense of crisis such as OS and Chip. However, risk together with opportunity that we needed to establish China ecosystem on AI.

Prof. Yang Qiang said we should think more about AI business model because we now had two critical factors that were computation power and cloud computing.

Prof. Zhou Ming said many cooperation with US citizen and experts. We did our best and we had many scenarios, data and technology. We can build up our internal cycle.

Prof. Chen Xiaoping said demand was not changed but the content of demand was changed. (e.g. what is next generation of home appliances?)

Finally, expert conclusions showed as follows.

Prof. Li Deyi said Next Generation of AI should be “Interactive”, “Can Learn”, “Able to Evolve”, “No Need Consciousness” (可交互、會學習、能進化、無意識). Consider AI as human intelligence extension!

Prof. Zhou Zhihua said open dynamic environment learning and integration of knowledge and logic reasoning would be the direction of machine learning. However, he said AI should be employable by industry.

Prof. Yang Qiang said we needed to overcome the gap between System 1 & System 2 and unity them.

Prof. Zhou Ming said AI did not become agents but AI learned their behaviors.

Prof. Tian Qi proposed to train up talents on AI+X.

Prof. Chen Xiaoping added rule change was AI’s opportunity.

Before the end of CCAI 2020, all panelists took a group photo for memory.