Quality Alchemist (品質煉金術師): CAAI Chinese Congress on Artificial Intelligence 2018 (中国人工智能大会)

2018年7月29日星期日

CAAI Chinese Congress on Artificial Intelligence 2018 (中国人工智能大会) - Day 2

The 4th Chinese Congress on Artificial Intelligence (CCAI2018) which was led by Chinese Association for Artificial Intelligence (CAAI), was held on July 28-29, 2018, in Shenzhen, China. As the largest official AI conference in China, CCAI has been held annually to promote the advancement of artificial intelligence globally. I attended the Congress and summarized it for sharing AI trend to all quality professionals.

During the congress, I met the Prof. Tie-niu Tan (谭铁牛) (Vice President of CAAI; Academician of CAS) and we took a photo for memory.

I also met Prof. KF Wong (黃錦輝) (Associate Dean (External Affairs), Faculty of Engineering, CUHK; Professor, SEEM Dept., CUHK) and we are both technical committee members in different discipline of CAAI. He appointed in the First Batch of Natural Language Processing Experts of CAAI. And I am Extenics Society committee members of CAAI since 2017.

Day 2 (29 July 2018):

In the beginning of the congress, Prof. Chengqing Zong (宗成庆) (Professor of CASIA) chair the morning reporting session.

Prof. Jiawei Han (韩家炜) (Abel Bliss Professor of UIUC; ACM/IEEE Fellow) was the first speaker and his presentation topic entitled “Mining Structures from Massive Text Data: A Cross Point of Data Mining, Machine Learning and Natural Language Processing”. Prof. Han said over 80% of our data is from text/natural language/social media, unstructured/semi-structured, noisy, dynamic, but inter-related! So that Prof. Han’s research roadmaps are “Mining hidden structures from text data”, “Turning text data into multidimensional text-cubes and typed networks” and “Mining cubes and networks to generate actionable knowledge”.

Firstly, Prof. Han introduced PathPredict studies to predict future co-authors of paper in the international research journal. Then he briefed the multi-dimensional text analysis. He also mentioned that the bottleneck was mining unstructured text for structures.

And then Prof. Han explained Phrase Mining concept from raw corpus to quality phrases and segmented corpus. He introduced how to judge the quality of phrases as follows:

i) Popularity (information retrieval vs cross-language information retrieval)

ii) Concordance

iii) Informativeness

iv) Completeness

After that Prof. Han explained Recognizing Typed Entities which enabling structured analysis of unstructured text corpus. However, Entity Recognition and Typing had 3 challenges and they were “Domain Restriction”, “Name ambiguity” and “Context Sparsity”. He used President Trump as example to brief the Fine-Grained Entity Typing. He also mentioned the Meta-Pattern Methodology in which separated Meta Patterns ($Country President $Politican), Entity (USA) and Attribute value (Barack Obama).

Lastly, Prof. Han stated a key problem from big data to big knowledge was multi-dimensional mining massive test data. And he shared his research journey to us.

Prof. Bing Liu (刘兵) (Distinguished Professor of UIC; AAAI/ACM/IEEE Fellow) was the second speaker and his topic named “Lifelong Learning, Continual Learning and Meta-Learning”. Prof. Liu told us about his drive license that even after passing the test, he continued to “learn on the job.” Over time, he got better and better that is real intelligence (not AlphaGo’s ability to play Go). Machine Learning (ML) is massive data-driven optimization but humans are not great at optimization.

Then Prof. Liu introduced the current ML paradigm for isolated single-task learning that needs a large number of training examples. ML is suitable for well-defined tasks in restricted and closed environment. But humans learning never learn in isolation or from scratch. Lifelong Learning (LL) is mimics this human learning capability. Therefore, the goat of LL is to create a machine that learns like humans.

The Lifelong Learning (LL) model was introduced as following diagram. The key characteristics of LL included:

i) Continual learning process

ii) Knowledge accumulation in Knowledge Base (KB) (long-term memory)

iii) Use and adapt the past learned knowledge to future learning, discovering new problems and learning them incrementally

iv) Learning on the job, learning in model application or testing (after the initial model built)

And then Prof Liu mentioned LL supervised and unsupervised learning included:

i) Lifelong naïve Bayesian classification

ii) Lifelong Topic Modeling (LTM)

iii) Lifelong learning in graph label propagation

The LL extended model was showed and he also briefed Meta-learning (also called learning to learn). At the end, he summarized that human learning was well-known but LL had a large space with huge challenges and it was still an open space with a lot of opportunities.

Prof. Jiaya Jia (贾佳亚) (Professor of CUHK; IEEE Fellow) was the third speaker and his presentation was “Advancement of Large-Scale Image Understanding and Segementation”. Firstly, he mentioned the importance of computer vision through daily life examples.

Then Prof. Jia demonstrated some research highlights included Deburring, Image Generation, Make-up Go, Face Edit, etc.

And then Prof. Jia discussed Semantic Segmentation included PSPNET (2017) and ICNet (2018); and Instance Segmentation included Multi-scale Patch Aggregation (2016), Sequential Grouping Networks (2017) and PANet (2018). He then introduced the end-to-end training.

After that Prof. Jia demonstrated the Instance Segmentation for auto-driving. He also introduced object modeling included Object Proposals, Aggregation of proposals Generalized Parts, Line Segments and Instance Boundaries. The first two modeling were Top-down methods and the last three modeling were Bottom-up Methods. At the end, Prof. Jia shared their excellent results on COCO competitions.

Prof. Zhi-Hua Zhou (Professor of Nanjing University; AAAI/ACM/IEEE Fellow) was the fourth speaker and his topic entitled “A Preliminary Exploration to Deep Forest”. Firstly Prof. Zhou introduced the AI, Machine Learning (ML) and Deep Learning (DL) and he said the wave of AI because of breakthrough in DL (especially on Images & Video, Speech & Audio, and Text & Language).

Then Prof. Zhou introduced Deep Learning (DL) quoted SIAM News (Jun 2017) that DL, a subfield of machine learning that uses “deep neural networks”, has achieved state-of-the-art results in fields such as image and text recognition. Deep Neural Network (DNN) was mentioned and employed Backpropagation (BP). BP is a supervised learning algorithm, for training Multi-layer Perceptrons (MLP) (Artificial Neural Networks).

And then Prof. Zhou briefed Deep Model. Most crucial for deep models were “Layer-by-layer processing”, “Feature transformation” and “Sufficient model complexity”. However, the disadvantage of DNN included too many parameters, model usually more complex than need, big training data, back box model and difficult for analysis.

After that Prof. Zhou introduced Deep Forest which was not Deep Neural Network (DNN) and no need for using BP algorithm for deep learning model. The gcForest is just a start of Deep Forest which used “Diversity” for ensembles.

Lastly, Prof. Zhou used Convolutional Neural Network (CNN) as example to explain new technology from starting to mature should be accumulated for long time. And he mentioned ZTE issues that we lacked of chip and algorithm (缺芯少魂). In order to reduce this risk, we needed to develop our own ML chip and system platform (avoid GPU and TensorFlow monopoly) and used Deep Forest (non-NN deep learning technology) to replace existing DNN.

In afternoon session, Prof. Bo Xu (徐波) (Director of Automation Institute of CAS; Dean of artificial intelligence institute of NUST) was session chair.

Prof. Mu-Ming Poo (蒲慕明) (Director of Institute of Neuroscience Academician of CAS, Academician of NAS) was the fifth speaker and his presentation title named “Brain Science and Brain-like Machine Learning”. In the beginning, Prof. Poo briefed the history of Brain Science.

Then Prof. Poo introduced the human neural network that contained more than 10^11 neural cells and 10^14 axon.

And then Prof. Poo briefed the three layer neural connection and they were macroscopic, mesoscopic and microscopic images. Mice 3D axon diagram was showed.

After that he demonstrate the zebra fish brain activities when escape. He explained Hebb’s Learning Rule (1949) "Cells that fire together wire together."

Finally, Prof. Poo concluded that we could benchmarking nature neural network for AI such as modulatory neurons, multi-directional connection, axon’s back-propagation and lateral propagation.

Forum – How to Advance the Development of Artificial Intelligence in China?

Ms. Quanling Zhang (张泉灵) (Founding partner of Znstartups) was session chair.

The first question Mr. Zhang asked for all guest about which one of three AI companies (AI chip, Autodrive car and ML solution) I should invest? It is difficult to answer from all guest that depended on situation.

Prof. Deyi Li (李德毅) (President of CAAI, Academician of CAE) shared his view on Autonomous Car that was a trend and implemented in China. In order to advance the development of AI in China, Prof. Li said we needed to established Smart Industry Alliance.

Then Prof. Bo Xu (徐波) (Director of Automation Institute of CAS; Dean of artificial intelligence institute of NUST) shared his view on this topic in automation research view.

Prof. Xu briefed the research planning in CAS included neural science, new generation AI development, complex system between human and machine, as well as AI+ industry. Lastly, he said “Stand at low tide and calm down at high tide” (低潮時堅守, 高潮時冷靜).

Dr. Haifeng Wang (王海峰) (Senior Vice President of Baidu; ACL Fellow) then shared AI Trend and Challenge. Technology challenges on basic theory were small size of sample, low energy consumption and explainable result. The challenges on application technology were perception technique and cognition technique. The challenges on industry were software and hardware integration, Deep Learning framework and AI Chip. The challenges on application systems were multi-technique integration use and system innovation in the integrated scenes. Lastly, he said AI system needed to continual evolution.