technology and ai sharing - from 2016 to y2017 and beyond
TRANSCRIPT
Technology and AI Sharing – From 2016 to Y2017 and Beyond
James CC Huang
最強大腦 – Human vs. Machine
Source: http://bit.ly/2jqiagE
最強大腦 – Human vs. Machine
Machine ( 小度 ) won!Source: http://bit.ly/2jqiagE
Share My 2016 Learning Journey
2016 台灣資料科學年會一天搞懂深度學習Jul 15-17給工程師的統計學及資料分析 123
Sep 4資料科學面面觀
Jan 23AWSome Day Express
Nov 22
AWS 基礎設施服務實作工作坊Dec 13
百度世界大會Sep 1
NVIDIA GTCSep 21-22
Deep Learning School
Sep 25-26
NIPSDec 5-10
CCAIAug 26-27
Put Things Together
Before We Start…• AL, machine learning, and deep learning are different, but in
the sharing we may not discuss about it.• Abbreviation:
– AI: Artificial Intelligence– ML: Machine Learning– DL: Deep Learning
• A lot of reference URL in the slides. Enjoy!– Articles / media reports / posts– Video clips
AI > Machine Learning > Deep Learning
Source: http://bit.ly/2h4AfLl
Best Short Definition of AI
Source: http://bit.ly/2h4z52B
AI = Training Data + Machine Learning + Human-in-the-loop
Technology Trend
Gartner's 2016 Hype Cycle for Emerging Technologies
* No “Deep Learning”
Source: link
Gartner’s Top 10 Strategic Technology Trends for 2017
Source: link, link
Top 10 Strategic Tech Trends - Intelligent AI & Advanced Machine Learning• AI, machine learning, deep learning, neural networks, natural language processing (NLP)• Parallel processing power + advanced algorithms + massive datasets• Real-time analytics
Intelligent Apps• Virtual personal assistants (VPAs)• Existing application with AI capabilities enabled.• 3 focus areas: advanced analytics, AI-powered and increasingly autonomous business
processes and AI-powered immersive, conversational and continuous interfaces.
Intelligent Things• Robots, drones, and autonomous vehicles.
Top 10 Strategic Tech Trends - Digital
Virtual & Augmented Reality• Training scenarios and remote experiences.• Enterprises should look for targeted applications of VR and AR through 2020.
Digital Twin• Dynamic software model + sensors• Users collaborate with data scientists and IT/BA professionals.
Blockchain• Bitcoin• FinTech
Top 10 Strategic Tech Trends - Mesh
Conversational Systems• Communicate across the digital device mesh (e.g., sensors, appliances, IoT systems) using text / voice / sight / sound /
tactile.
Mesh App and Service Architecture (MASA)• Flexible enough to allow rapid evolution of user needs and how they interact with technology.• Apps connect and communicate and with other apps using agile architecture with, for example, HTTP/REST JSON.
Digital Technology Platforms• Information systems, customer experience, analytics and intelligence, IoT and business ecosystems.• New platforms and services for IoT, AI and conversational systems will be a key focus through 2020.
Adaptive Security Architecture• Multilayered security and use of user and entity behavior analytics will become a requirement for virtually every
enterprise.• Security in the IoT environment
With data, advanced AI, and computing power, everything will be “more” intelligent.
Programming Language and Tool RankingFOCUSING ON DATA SCIENCE AND AI / MACHINE LEARNING / DEEP LEARNING
Top Programming Language - KDnuggets
Source: link
Top Data Science Tools - KDnuggets
Source: link
Top 20 Python ML Open Source Project
Top projects are ML, DLProjects on GitHub. A lot of them are new in top 20 in Y2016.
Source: link
DL Software w/ Default Support for AWS and PythonSoftware Platform Interface GPU
SupportRecurrent nets
Convolutional nets
RBM/DBNs Parallel execution
Caffe
Linux, Mac OS X, AWS, Windows support by Microsoft Research
C++, command line, Python, MATLAB Yes Yes Yes No Yes
Deeplearning4j
Linux, Mac OS X, Windows, Android (Cross-platform)
Java, Scala, Clojure Yes Yes Yes Yes Yes
KerasLinux, Mac OS X, Windows
PythonYes Yes Yes Yes Yes
Microsoft Cognitive Toolkit - CNTK
Windows, Linux (OSX via Docker on roadmap)
Python, C++, Command line, BrainScript (.NET on roadmap) Yes Yes Yes No Yes
MXNetLinux, Mac OS X, Windows, AWS, Android, iOS, JavaScript
C++, Python, Julia, Matlab, JavaScript, Go, R, Scala
Yes Yes Yes Yes Yes
PaddlePaddle Linux, Mac OS X Python, C++ Yes Yes Yes ? Yes
TensorFlowLinux, Mac OS X, Windows Python, (C/C++ public API only for
executing graphs) Yes Yes Yes Yes Yes
Theano Cross-platform Python Yes Yes Yes Yes Yes
TorchLinux, Mac OS X, Windows, Android, iOS
Lua, LuaJIT, C, utility library for C++/OpenCL Yes Yes Yes Yes Yes
Source: link
Evaluate
Which is the best programming language to data / AI / ML / DL?
How to select deep learning software?
On-premise or cloud / API platform?
Use Case:Eva can get current product customer account on Facebook Messenger chatbot using natural language query and voice command.
Microsoft Cognitive Services
[Video] Microsoft Cognitive Services: Introducing the Seeing AI projecthttp://bit.ly/2i8JOgc
Artificial Intelligence, Machine Learning, Deep Learning
AI Talent Wars / Acquisition• Giant corporations are soaking up AI talent.• Top AI researchers -> industry with humongous data.• “The cost of acquiring a top AI researcher is comparable to
the cost of acquiring an NFL quarterback.”• AI talent shortage.
Source: link, link
Academic researcher -> Humongous data and computing power
For AI talent, hire from outside, or train and transit our developers for AI-powered projects?
Gap for the Transition• Academic background• Differences between computer program and brain (AI tries
to simulate brain)– Computer program: define the general to store specifics– Brain: store the specific to identify the general
Source: link
Rise of Current AI (Not Long Time Ago)
AI > Machine Learning > Deep Learning
Source: http://bit.ly/2h4AfLl
One of the Biggest Crowdsourcing Project
– Started in Y2007– On Amazon Mechanical Turk Marketplace
• 48,940 workers• 167 countries
– Total number of images: 14,197,122 (as of 2010/4/30)
ImageNet Challenge
ILSVRC’16 winner: Error rate 2.991%
Super-human precision
* Human-level performance: 5.1%
2016: The Year That Deep Learning Took Over
Notable AI Events in 2016 (by China)
Source: link
Google Trends - Deep Learning
Published AI Documents by Country (Y2015, Top 10)
* Taiwan ranked #11.Source: link, link
Main Developments in 2016 (From Top AI Researchers)
Reinforcement Learning
Inhuman Encryption GAN NLP
Machine Translation Lip Reading Speech
Recognition WaveNet
Computer Vision Hype
Source: link
AI Category Innovation Quadrant
Source: http://bit.ly/2h4FA5CSource: link
DL dominates now (and is still growing fast).
* DL is not equal to AI / ML.
Neural Networks Zoo
Most talked: CNN (?)Hot: RNN (?)Uprising: GAN (?)
Source: link
Rule of Thumb (Mostly from Andrew Ng)Why
• Add value to our business.
When• “If a typical person can do a mental task with less than one second of thought, we can probably automate it
using AI either now or in the near future.”
What• A large amount of data.
How• Choose tool(s) and “customize to our business context and data.”
Evaluation• If AI error rate surpasses human-level performance.
Source: link
Predictions for AI in 2017
5 Big Predictions for AI in 2017 (MIT Press)Positive reinforcement
• Reinforcement Learning• AlphaGo -> Master -> ?
Dueling neural networks• GAN (Generative Adversarial Networks)• Learn from unlabeled data
China’s AI boom
Language learning• NLP• Image caption -> description
Backlash to the hype
Source: link
Key Trends in 2017 (From Top AI Researchers)
NLP Unsupervised Learning
Deep Learning in Healthcare Chatbot
Self-driving Car Computer VisionHybrid deep
learning with other ML/AI techniques
AutoML
Commodify Deep Learning
Source: link
High Performance Computing (HPC)BOOST AI / ML / DL
In the race to build the best AI, there’s already one clear winner
中國大陸人稱“皮衣教主”Source: link
GTC 2016 (GPU Technology Conference)AI Revolution
GPU Supercomputer & Acceleration for Data Center
Computer Vision, VR
AI City by Y2020 (1B+ Cameras)
Self-Driving Car
AI Computing Ecosystem
Source: link
NVIDIA DGX-1 vs. Supercomputers
Unit: teraflops
Intel Core i7-6700HQ
DGX-1
China (神威 ·太湖之光 )
Japan (Future)
0 20000 40000 60000 80000 100000 120000 140000
0.03325
170
93000
130000
* DGX-1 list price: US$ 130,000
[Video] GPU vs. CPU on training MNIST dataset
HPC Competition (On-going)
CPU + GPU?CPU + FPGA?Tailored Processor?
HPC Competition (On-going)• GPU is current leader.• Major cloud computing platforms support both GPU and
FPGA, e.g.
Major DL Software Supports GPU AccelerationSoftware Platform Interface GPU
SupportRecurrent nets
Convolutional nets
RBM/DBNs Parallel execution
Caffe
Linux, Mac OS X, AWS, Windows support by Microsoft Research
C++, command line, Python, MATLAB Yes Yes Yes No Yes
Deeplearning4j
Linux, Mac OS X, Windows, Android (Cross-platform)
Java, Scala, Clojure Yes Yes Yes Yes Yes
KerasLinux, Mac OS X, Windows
PythonYes Yes Yes Yes Yes
Microsoft Cognitive Toolkit - CNTK
Windows, Linux (OSX via Docker on roadmap)
Python, C++, Command line, BrainScript (.NET on roadmap) Yes Yes Yes No Yes
MXNet Linux, Mac OS X, Windows, AWS, Android, iOS, JavaScript
C++, Python, Julia, Matlab, JavaScript, Go, R, Scala
Yes Yes Yes Yes Yes
PaddlePaddle Linux, Mac OS X Python, C++ Yes Yes Yes ? Yes
TensorFlowLinux, Mac OS X, Windows Python, (C/C++ public API only for
executing graphs) Yes Yes Yes Yes Yes
Theano Cross-platform Python Yes Yes Yes Yes Yes
TorchLinux, Mac OS X, Windows, Android, iOS
Lua, LuaJIT, C, utility library for C++/OpenCL Yes Yes Yes Yes Yes
Source: link
Future Challenges and Opportunities
2016: Rise of AI2017: AI Enabled Things
AI Winter? Bubble?AI is fueled by• Humongous data (or Big Data)• Algorithms• Hardware advances• Entry barrier lowers (*)
Source: link
Hype? Fake AI?
Source: link
Media Hype Academic
and Research
Ultimate AI?
Value-added Business with AI/ML/DL
Use Case: Real-time Answer for TS / SalesSource: link
How to Add Value to OUR Business with AI
DIY AI
Source: link
DIY AI“Cat Shooting” system• Cat triggers camera.• Pop up sprinkler.• “Water”!
NVIDIA Jetson TX1 list price: US$599
Source: link
Is “Current” AI Smart?1. Ask Allo “What should be my New Year’sresolution be?” Ask several times to getmore resolutions.
2. See what you get!
3. Did you get the same answers in thearticle?
4. Is this the AI we look forward to?Source: link
Source: link
Job Displacement
Source: link
Source: link
Cybersecurity• Malicious tool applying AI• Defense leveraging AI• Fake data to influence result and output
Source: link, link
“Anatomy of the Blockbuster Novel”• #NLP #MachineLearning
#TextMining• Topic Modeling• Sentiment Analysis• Writing styles
– Frequently used words– Punctuation
Source: link
What’s Next