從 Project Theta 到台灣人工智慧學校

From Project θ to Taiwan AI Academy

Sheng-Wei ChenResearch Fellow, Academia Sinica

Chairman, Taiwan Data Science Foundation

Change is the only constant- Heraclitus (535 BC - 475 BC)

陳昇瑋 / 從大數據走向人工智慧 6


Mastering Chess and Shogi by self play with reinforcement learning

陳昇瑋 / 從大數據走向人工智慧

AI IN MANUFACTURING

8


2016 Global manufacturing competitiveness index rankings


1/3 of the GDP

Manufacturing GDP of $178B, almost 1/3 of total GDP

30% of the employment are in the manufacturing sector

Cheap labor cost of $9.42/hr with average labor productivity of almost $60k in GDP/person

17% corporate tax rate

10


McKinsey’s Four Dimensions in AI Value Chain

11

Smart R&D and

forecasting

ProjectOptimized

production with

lower cost and

higher efficiency

ProduceProducts and

services at the

right price, time,

and targets

PromoteEnriched and

tailored user

experience

Provide


The Four-P Dimensions in Manufacturing Improve product design Automate supplier assessment and price negotiation Anticipate parts requirements

Improve manufacturing processes Automate assembly lines limit product rework

Optimize pricing Predict sales of maintenance services Refine sales-leads prioritization

Optimize flight/fleet planning and route Enhance maintenance engineering Enhance pilot training

12

Provide

Project

Promote

Produce

14

Professor

• Prior to joining Harvard in 1992, Dr. Kung taught at Carnegie Mellon University for 19 years.

• In 1999 he started a joint Ph.D. program with colleagues at the Harvard Business School on information, technology, and management, and co-chaired this Harvard program from 1999 to 2006.

• Member of National Academy of Engineering• Guggenheim Fellowship• IEEE Computer Society Charles Babbage Award

HT Kung

• Academician, Academia• William H. Gates Professor, Harvard John A.

Paulson School of Engineering and Applied Sciences

Current

Past Experiences

16

To empower Taiwan (manufacturing) industries with Artificial Intelligence

February – November in 2017

17

台塑石化長春石化奇美實業英業達欣興電子敬鵬工業可成科技致茂電子永進機械研華科技農科院紡織所聯發科技台積電宏遠紡織台元紡織佳和紡織強盛染整臺灣塑膠龍鼎蘭花經緯航太科技

Unmet “Soft” Needs for Nurturing Next-Generation Industries in the AI Era

Human resource development Machine learning experts with hands-on experiences

Problem/opportunity identificationProblem identification is the biggest challenge for newcomers

Business transformationProblem identification and solving strategiesSpin-offs/R&D initiatives

Shared technology infrastructureKnowledge base, datasets and baseline practices

18

人工智慧發展策略建議

Project θ Weekly Meetings at NCTUand Also Online with Academia Sinica

June 6, 2017


Why it’s called Project θ?

20


Industry-wide Problems

Automated Optical Inspection (AOI) systems

Adaptive process control

Predictive maintenance

Component selection optimization

…

21


A sad story that AI-assisted AOI can help avoid 14 suicide events in 2010 at Foxconn China factories

Only 2 of the suicides survived

Industry-wide Problem #1:Human Operators for Optical Inspection

https://theinitium.com/article/20170802-mainland-Foxconn-factorygirl/



Human Operators for Optical Inspection

The factories recruit only workers under 29 years old

Their work involve checking scratches on consumer products (likely Apple iPhone) for 2,880 times a day

This means 4 times per minute assuming 12 working hours per day




Typical metal defects

24

人工智慧發展策略建議 25

Typical PCB defects


Typical defects after SMT (Surface-Mount Technology) process短路

空焊

極反

缺件

浮高

跪腳

撞件

錫球

墓碑

…

26

https://www.researchmfg.com/2011/02/soldering-defect-symptom/

https://www.researchmfg.com/2011/02/soldering-defect-symptom/

More SMT/DIP Defect Examples

27

Typical Deep-Learning based AOI Systems

30

AI

…

OK

Deep Neural Networks

Deep Convolution Neural Networks

Transfer learningPre-trained using 14-million image dataset

Resnet with > 8-million parameters

Input images Model training / inference

OK

OK

Case study –Human vs. Neural Inspection

33

4 human inspectors for 23 product linesThroughput: 300K patches per human per day = 1.2M patches per dayLeakage rate between 5% to 10% while False alarm rate > 10%

Human

AI

Equipment: A PC with NVIDIA GeForce 1080 Ti(4,000 USD)Throughput: 167 patches per second = 10 K patches per minute = 14M patches per dayLeakage rate < 0.01% while False alarm rate < 5%


Industry-wide Problem #2:Adaptive process control

35


Case Study: A Chemical Process

12 parameters

Hydrogen (H)

Catalyst

Ethylene (C2H4), Ethane (C2H6), Butene (C4H8)

Pressure, temperature, fluid level, and so on

Output

A quality index of a certain chemical product

36

Acceptable range

Yield rate: 61%

Qua

lity

Inde

x

Residual networks

Very similar to Residual network in Image classification

main stream + residuals

38

Residual network reference

Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks, Pranav et.al., 2017

39

Preliminary Control Results

Human yield rate: 61%CNN yield rate: 98%

Qua

lity

Inde

x


Industry-wide Problem #3:Predictive maintenance

Especially important for equipment with high failure cost (such as motors in machine tools)

Also important for expensive consumables (such as blades used in precision cutting machines)

40


Industry-wide Problem #4:Component selection optimization

41


te

42

Model & workflow

Pigment selection

model31點反射率

32個染料 (0, 1)

•Multi-classification (每筆observation 最多 3 個 1)

Pigment concentration

model

31點反射率＊ 4 組 , 共124 dim

•一組是該顏色的反射率

•其他 3組是染料對應的反射率(固定一種濃度, 1.5%)

3個染料濃度 (0 –5 %)•Multiple output

regression

Reflectance prediction

model

32個染料濃度(沒有用則 0)

31點反射率 (0 –1)•Multiple output

regression

43

Input Output


Pigment 1 Pigment 2 Pigment 3


PROJECT Θ TEAM HAS SOLVED

10+ PROBLEMSFROM 10+ COMPANIES

WITHIN 6 MONTHS…

45


LOOKS IT WORKS OUT, BUT …

46


Challenges for AI Development in Taiwan

Wide gap between academia and industry

Lack of experienced talents Used to adopt rather than develop

technology

TAIWAN AI ACADEMY－A SOLUTION TO SCALE OUT PROJECT Θ

68

http://aiacademy.tw/

Address the “lack of AI talents” problem

Offer short, intensive and scalable training courses

Aim to train >= 1500 talents each year



Domain experts + AI

Strong linkage with the academy

Real-life problems from industry as exercises and term projects


72

Corporate Partner ProgramCorporates provide real-life problems (and datasets)Students tackle these problems as term projectsCorporates may recruit students after they finish the training courses

Current class design

Elite Engineer Class (技術領袖培訓班)12 weeks9am to 6pm on Monday to FridayLectures + hands-on sessions + term projectsMid-term and final exams

Manager Class (經理人周末研修班)12 weeks9am to 9pm each Saturday Lectures only

73

Elite Engineer Class

74

Applications due on Dec 4, 2017. Nearly 500 applicants registered while we can only accept 208 students.

Two-step filtering:1. Document review2. Entrance exam: calculus, linear

algebra, probability, statistics, programming

Empowering Taiwan in the AI Era!Sheng-Wei ChenAcademia Sinica

從 Project Theta 到台灣人工智慧學校

Education