9. 陳昇瑋 / 台灣產業 AI 化 - 如何跨出第一步?
機器學習
24
Training a prediction machine by
showing examples instead of
programming it.
-Yann LeCun
(prediction machine: 可基於已知預測未知的數學模型)
10. 陳昇瑋 / 台灣產業 AI 化 - 如何跨出第一步?
機器學習的定義
25
讓電腦能從資料裡頭淬取出
規則的演算法。
Find the common patterns
from the left waveforms
It seems impossible to
write a program for
speech recognition
你好 你好
你好 你好
You quickly get lost in the
exceptions and special cases.
(Slide Credit: Hung-Yi Lee)
11. 陳昇瑋 / 台灣產業 AI 化 - 如何跨出第一步?
就放棄教電腦規則,讓它自己學吧!
你好
大家好
人帥真好
You said
“你好”
很多訓練資料
機器學習演算法
從訓練資料中
找到規則
(Slide Credit: Hung-Yi Lee)
13. 陳昇瑋 / 台灣產業 AI 化 - 如何跨出第一步?
現代機器學習算法所學到的複雜規則
29
14. 陳昇瑋 / 人工智慧民主化在台灣
A Gaydar
30
Based on 35,000 facial images
Human judges: 61% for men, 54% for women
AI judges: 91% for men, 83% for women
A heat map of where the algorithm looks to detect signs of homosexuality (Kosinski and Wang)
https://osf.io/zn79k/
58. 陳昇瑋 / 人工智慧民主化在台灣
Natural Language Processing (NLP): Some
Sentence Generation Examples by GTP-2
103
Bad news:
The technology can be used in generating fake news
Transformer based
(Slide credit: HT Kung)
59. 陳昇瑋 / 人工智慧民主化在台灣
More text generation samples
104https://openai.com/blog/better-language-models/
60. 陳昇瑋 / 人工智慧民主化在台灣
BERT and GPT-2
105
● Both BERT (from Google) and GPT-2 (OpenAI) are general purpose
pretrained NLP Feature Extractors based on the Transformer trained on
enormous amounts of text data.
● These models can be fine-tuned on small-data NLP tasks (like question
answering), resulting in substantial accuracy improvements compared to
training on these datasets from scratch
Pretrained BERT/GPT-2
Additional classifier for
your own tasks
The pretrained feature
extractor
Based on
Transformer
61. 陳昇瑋 / 人工智慧民主化在台灣
Advances of NLP Models in Recent Years
106
Deeper models, Larger Datasets
Better Features!
Performance Improvement: ~20%
Reuse of pre-training BERT/GPT-2: models.
Impact is similar to that of ImageNet in computer vision
Paper on theTransformer:Vaswani, A. et al, “Attention Is AllYou Need,” NIPS 2017
One-hot
Word
Embedding
ELMo BERT GPT-2
milestone
Transformer
based
BooksCorpus
(800M words) +
Wikipedia
(2,500M words)
scraped content from
the Internet of 8
million web pages
WMT 2011
(800M words)
WMT 2011
(800M words)
Large-scale dataset like ImageNet
RNN
based
Development of pretrained
‘feature extractor’ for NLP tasks:
68. 陳昇瑋 / 人工智慧民主化在台灣
Typical Applications of RL
Play games: Atari, poker, Go, ...
Explore worlds: 3D worlds, Labyrinth, ...
Control physical systems: manipulate, walk, swim, ...
Interact with users: recommend, optimize, personalize,
...
121
(Slide credit: David Silver)
69. 陳昇瑋 / 人工智慧民主化在台灣
More RL Applications
Flying Helicopter
Driving
Google Cuts Its Giant Electricity Bill With DeepMind-
Powered AI
Parameter tuning in manufacturing lines
Text generation
Hongyu Guo, “Generating Text with Deep Reinforcement
Learning”, NIPS, 2015
Marc'AurelioRanzato,SumitChopra,Michael Auli,Wojciech
Zaremba, “Sequence Level Training with Recurrent Neural
Networks”, ICLR, 2016
122(Slide Credit: Hung-Yi Lee)
74. 0.95
F-score
Algorithm Ophthalmologist
(median)
0.91
“The study by Gulshan and colleagues truly
represents the brave new world in
medicine.”
“Google just published this paper in JAMA
(impact factor 44.405) [...] It actually lives
up to the hype.”
Dr. Andrew Beam, Dr. Isaac Kohane
Harvard Medical School
Dr. Luke Oakden-Rayner
University of Adelaide
75. 陳昇瑋 / 人工智慧民主化在台灣
Deep Learning for Detection of Diabetic
Eye Disease (2016)
132
Algorithm’s F1-score: 0.95
Median F1-score of 8 ophthalmologists : 0.91
82. 陳昇瑋 / 人工智慧民主化在台灣 143
Deep Learning for Kidney Function Classification and
Prediction using Ultrasound-based Imaging
Chin-Chi Kuo1, Chun-Min Chang2, Kuan-Ting Liu2, Wei-Kai Lin2,
Chih-Wei Chung1, and Kuan-Ta Chen2
1Big Data Center, China Medical University Hospital, China Medical University, Taichung, Taiwan
2Institute of Information Science, Academia Sinica, Taiwan
eGFR
(腎功能指數)
84. Cardiologist-Level Arrhythmia Detection
with Convolutional Neural Networks
150
Goal: diagnose irregular heart rhythms, also known as
arrhythmias, from single-lead ECG signals better than a
cardiologist
85. 陳昇瑋 / 人工智慧民主化在台灣
Input and Output
Input: a time-series of raw ECG signal
The 30 second long ECG signal is sampled at 200 Hz
From 29,163 patients
Output: a sequence of rhythm classes
The model outputs a new prediction once every second
Total 14 rhythm classes are identified
151
87. 陳昇瑋 / 人工智慧民主化在台灣
Model
34 layers NN
16 residual blocks
2 conv layers per block
Filter length = 16 samples
# filter = 64*k, k start from 1 and is
incremented every 4-th residual block
Every residual block subsamples
its input by a factor of 2
153
90. Predictive tasks for healthcare
Given a large corpus of training data of de-identified medical records, can we
predict interesting aspects of the future for a patient not in the training set?
● will patient be readmitted to hospital in next N days?
● what is the likely length of hospital stay for patient checking in?
● what are the most likely diagnoses for the patient right now? and why?
● what medications should a doctor consider prescribing?
● what tests should be considered for this patient?
● which patients are at highest risk for X in next month?
Collaborating with several healthcare organizations, including UCSF, Stanford, and
Univ. of Chicago. Have early promising results (no public paper yet)
94. A spectrum of learning tasks
• Low-dimensional data (e.g. less
than 100 dimensions)
• Lots of noise in the data
• There is not much structure in
the data, and what structure
there is, can be represented by a
fairly simple model.
• The main problem is
distinguishing true structure
from noise.
• High-dimensional data (e.g.
more than 1000+ dimensions)
• The noise is not sufficient to
obscure the structure in the data
if we process it right.
• There is a huge amount of
structure in the data, but the
structure is too complicated to
be represented by a simple
model.
• The main problem is figuring out
a way to represent complicated
structure that allows it to be
learned.
Statistics --------------------- Deep Learning
99. 陳昇瑋 / 人工智慧民主化在台灣
Algorithmic Trading
176
Algorithmic finance is roughly a $1 trillion market (as of early 2018)
Growing at 10.3% CAGR [1]
90% volume in the public equities markets
50% volume in the futures markets
6 of the top 10 hedge funds based on performance are quant funds and most of
those are quant funds in commodities markets.
[1] source:Technavio’s 2016 report: Global AlgorithmicTrading Market 2016-2020
103. 陳昇瑋 / 人工智慧民主化在台灣
But, AI is still far from perfect
180
http://www.eurekahedge.com/Indices/IndexView/Eurekahedge/683/Eurekahedge_AI
_Hedge_fund_Index
104. 陳昇瑋 / 人工智慧民主化在台灣
Correlation between AI funds and humans
has risen in past years
181
https://www.bloomberg.com/news/articles/2018-03-12/robot-takeover-
stalls-in-worst-slump-for-ai-funds-on-record
106. 陳昇瑋 / 人工智慧民主化在台灣
Common use of machine learning
Prediction of direction
Prediction of price
Prediction of risk
Prediction of portfolio performance
Detection of big players’ movements
Rebalancing
Liquidating
…
Fully automation of trading
183
https://www.quora.com/Is-there-any-hedge-fund-using-
machine-learning-based-algorithms-for-trading
108. 陳昇瑋 / 人工智慧民主化在台灣
Alternative data and its applications
Individuals
Twitter sentiment data to trade the broad equity market (iSentium)
News sentiment data to trade bonds, equities and currencies (RavenPack)
Web search, personal email, app downloads, job listing, …
Business processes
Consumer transaction data to trade individual stocks (Eagle Alpha)
PoS, payments, invoices / receipts, real estates, cargo, retail prices, phone
calls, shipments via logistics, lodging, taxi, insurance, …
Sensors
Geolocation data to estimate retail activity and trade individual stocks
(Advan Research)
Satellite image data to estimate employee activity and trade individual
stocks (RS Metrics)
185
109. 陳昇瑋 / 人工智慧民主化在台灣
iSentium: Tweets
186
Construction of iSentium Daily Directional Indicator
The universe is limited to the 100 stocks which are most representative of the S&P 500,
filtered using tweet volume and realized volatility measures.
Tweets are assigned a sentiment score using a patented NLP algorithm.
By aggregating tweet scores, a sentiment level is produced per minute between 8:30 AM and
4:30 PM every day. Sentiment for the day is aggregated using an exponentially weighted
moving average over the past ten days.
S&P 500 returns are forecasted using a linear regression over the sentiment scores for the
past two days, with betas evolved via a Kalman filter.
110. Eagle-Alpha: Email Receipts
187
We have analyzed a dataset of email receipts for 97 companies. 36 of these were private
companies, and 61 public, 31 of which were S&P 500 constituents. Taking liquidity into
consideration, we decided to test trading signals for the S&P 500 companies only.
114. 陳昇瑋 / 人工智慧民主化在台灣 225
Mobile computing, inexpensive sensors collecting terabytes of data, and the
rise of machine learning that can use that data will fundamentally change the
way the global economy is organized.
- Fortune, “CEOs:The Revolution is Coming,” March 2016
133. Convolution Neural Networks + Transfer Learning
Pre-trained using 14-million image dataset
ResNet with > 8-million parameters
Input
images
Model training /
inference
OK
OK
以深度學習進行自動瑕疵檢測
139. 台灣人工智慧學校首屆開學典禮
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)
257
產業共通挑戰 #3-預測性維護
193. 陳昇瑋 / 人工智慧民主化在台灣
Execute pilot projects to gain momentum
Build an in-house AI team
Provide broad AI training
Develop an AI strategy
Develop internal and external communications
341
194. 陳昇瑋 / 人工智慧民主化在台灣
State of AI In The Enterprise, 2018
347
Deloitte interviewed 1,100 IT and line-of-business executives
from US-based companies in the 3rd quarter of 2018.
82% of enterprise AI early adopters are seeing a positive ROI from
their production-level projects this year.
69% of enterprises are facing a “moderate, major or extreme”
skills gap in finding skilled associates to staff their new AI-driven
business models and projects.
63% of enterprises have adopted machine learning, making this
category the most popular of all AI technologies in 2018.
205. 陳昇瑋 / 從大數據走向人工智慧
持續的團隊支援
366
A common data platform and workflow is
crucial for enterprise success.
Data Engineer ML Engineer Biz Analyst DevOps DevOps +
ML Engineer
App
Developer
(Credit: IBM Systems Lab Services)
(all under the supervision of Data Scientist)
206. 陳昇瑋 / 從大數據走向人工智慧
AI vs. BI
AI systems suggest decisions for users by making
predictions
BI systems support users make decisions based on data
visualization
Key difference
AI systems are based on generalizable models
BI systems require humans to generalize
Best practice
AI: fast, massive, error-tolerant, ML-capable problems
BI: otherwise
AI+BI: making sense of AI decisions
369
210. 陳昇瑋 / 人工智慧民主化在台灣
What we can and cannot today
What we can have
Safer car, autonomous car
Better medical image
analysis
Personalized medicine
Adequate language
translation
Useful but stupid chatbots
Information search,
retrieval, filtering
Numerous applications in
energy, finance,
manufacturing,
commerce, law, …
What we cannot have
(yet)
Machine with common
sense
Intelligent personal
assistants
“Smart” chatbots
Household robots
Agile and dexterous robots
Artificial General
Intelligence (AGI)
392(Credit:Yann LeCun)
211. 陳昇瑋 / 人工智慧民主化在台灣
Strong AI Weak AI
Can think
Own conscious
Act as it can think
Consciousless
(1980)
216. 陳昇瑋 / 人工智慧民主化在台灣
AI Don’t Know What They are Talking About
404
https://www.facebook.com/playgroundenglish/videos/629372370729430/?hc_ref=ARQ
HCaS2GZ9jUgZermEupF5yerADq2X9F9P40OR3n70poUiCy7R0X3oHrGxyLSrWVdI
217. Change is the only constant.
- Heraclitus (535 BC - 475 BC)
(Slide Credit: Albert Chen)
224. 陳昇瑋 / 人工智慧民主化在台灣 421
(Source: Future of Jobs Report 2018, World Economic Forum)
225. 陳昇瑋 / 人工智慧民主化在台灣
AI outperformed 20 corporate lawyers at
legal work
422
Challenge: review risks contained in five non-disclosure agreements (NDAs).
AI vs. associates and in-house lawyers from global firms such as Goldman Sachs, Cisco
and Alston & Bird, as well as general counsel and sole practitioners.
AI matched the top-performing lawyer for accuracy – both achieved 94%. Collectively,
the lawyers managed an average of 85%, with the worst performer recording 67%.
AI: 26 seconds; lawyers’ average: 92 minutes, where the speediest lawyer took 51
minutes
https://www.weforum.org/agenda/2018/11/this-ai-outperformed-20-corporate-
lawyers-at-legal-work/
226. 陳昇瑋 / 人工智慧民主化在台灣
The Jobs Landscape in 2022
423(Source: Future of Jobs Report 2018,World Economic Forum)
Note: Roles marked with * appear across multiple columns. This reflects the fact that they might be seeing stable or declining demand across one
industry but be in demand in another.
*
*
227. 陳昇瑋 / 人工智慧民主化在台灣
人機協作
Camelyon Grand Challenge in 2016
根據切片檢查偵測轉移性乳癌
The Winning Team
深度學習演算法: 92.5%
病理科醫師: 96.6%
兩者合作: 99.5%
人類與機器擅長不同的預測層面
人類與機器犯不同類型的錯。
確認這兩種不同的能力,結合人類與機器的預測來克服這
些弱點,這樣的組合可以大幅減少錯誤率。
428
228. 陳昇瑋 / 人工智慧民主化在台灣
machine-learning model from 30,000+ deals from the last decade that draws from
many sources, including Crunchbase, Mattermark, and PitchBook Data. For each deal,
we looked at whether a team made it to a series-A round by exploring 400 features and
identified 20 features as most predictive of future success.
One of the insights we uncovered is that start-ups that failed to advance to series A had
an average seed investment of $0.5 million, and the average investment for start-ups
that advanced to series A was $1.5 million.
Another example insight came from analyzing the background of founders, which
suggests that a deal with two founders from different universities is twice as likely to
succeed as those with founders from the same university.
from the 2015 cohort of seed-stage companies, 16 percent of all seed-stage companies
backed by VCs went on to raise series-A funding within 15 months. By comparison, 40
percent of recommended by ML (2.5 times improvement)
Human + AI would yield the best performance: 3.5 times the industry average
429
https://www.mckinsey.com/industries/high-tech/our-insights/a-machine-learning-
approach-to-venture-capital
231. 陳昇瑋 / 人工智慧民主化在台灣
Take-Home Messages
今天的 AI 如同 1994 年的 World Wide Web
今天學 ML 可能如同 1994 年學寫 HTML and CGI
技術快速進展及堆疊,這是最值得投資技術的時代
AI 在許多領域有殺手級應用,是技術人投入重點領域的黃
金時代
AI 不會在這裡停住,AI 技術才剛開始發展
442