The Future of AI, Nowmsd2018.metrodata.co.id/microsite-2015/images/key...lupakanlah daku tinggalkan...
-
Upload
trinhthien -
Category
Documents
-
view
219 -
download
0
Transcript of The Future of AI, Nowmsd2018.metrodata.co.id/microsite-2015/images/key...lupakanlah daku tinggalkan...
75% of
commercial
enterprise apps
will use AI by 2020
In 2018, blended AI will
disrupt your customer
service and sales strategy
85% of CIOs
will be piloting AI
programs by 2020
Knowledge workers spend
80% of their time
searching and preparing
data... NOT on innovating
with Data Science and AI
2017 IDC report: http://idcdocserv.com/US42072917e_IBM
Innovation Trigger: A potential technology breakthrough kicks things off. Early proof-of-concept stories and media interest trigger significant publicity. Often no usable products exist and commercial viability is unproven.
Peak of Inflated Expectations: Early publicity produces a number of success stories — often accompanied by scores of failures. Some companies take action; many do not.
Trough of Disillusionment: Interest wanes as experiments and implementations fail to deliver. Producers of the technology shake out or fail. Investments continue only if the surviving providers improve their products to the satisfaction of early adopters.
Slope of Enlightenment: More instances of how the technology can benefit the enterprise start to crystallize and become more widely understood. Second- and third-generation products appear from technology providers. More enterprises fund pilots; conservative companies remain cautious.
Plateau of Productivity: Mainstream adoption starts to take off. Criteria for assessing provider viability are more clearly defined. The technology's broad market applicability and relevance are clearly paying off.
Gartner Hypecyle for Emerging Technology July 2018: IoT & AI Platform
About10-Years
Name: SophiaWho am I?: Social humanoid
robotMy parent?: Hong Kong-
based company Hanson Robotics
Born: Activated on April 19, 2015
Shown to public: SXSW festival, March 2016 (USA)
Capabilities: Display 62+ facial expressions
Data Science Artificial Intelligence: Machine Learning
Regression
Clustering
Self Driving Car
Autonomous Vehicle
Learn to win by Observing
Watson IoT Platform
Source: SWG Insight, Edisi Q3 2017
Watson IoT Edge Analytics Watson Realtime Insights Watson Machine Learning Blockchain Spatial Analytics (Automatic routing given
GPS coordinates) The Weather Channel
IoT Devices
Industry 4.0 - Analyze data stream from IoT devices e.g. Supply Chain AutomationIBM Watson IoT Platform with Blockchain and Machine Learning
A group of young people standing around a party
An
dro
id M
ob
ile A
pp
by
An
dre
w W
idja
ja
Ba
se m
od
el: h
ttp
s://
git
hu
b.c
om
/ben
iz/d
eep
det
ect
Cross Industry – (Image Captioning) Translate Video/Image to meaningful textTrigger selected commercial (iklan) for exampleIBM Watson Visual Recognition, Watson Studio
A Street with a sign for Cars and People
An
dro
id M
ob
ile A
pp
by
An
dre
w W
idja
ja
Ba
se m
od
el: h
ttp
s://
git
hu
b.c
om
/ben
iz/d
eep
det
ect
Cross Industry – (Image Captioning) Translate Video/Image to meaningful textTrigger selected commercial (iklan) for exampleIBM Watson Visual Recognition, Watson Studio
IoT-Enabled device running Machine Learning model “Face Recognition” locally Embedded Computing NVidia Jetson TX2 Supercomputer
7.5 Watt Hardware 256 GPU cores, 8GB RAM
Neural Network Model can be trained by Watson Studio Programmed in Python State-of-the-art Face Recognition
Library with dlib
WatsonStudio
Cross Industry – Realtime Face RecognitionWatson Studio deployed to Edge IoT Device
Source http://www.thejakartapost.com/travel/2018/08/04/jakpost-guide-to-2018-asian-games-west-java.html
Cross Industry – Classify/Understand Natural Language as humans doIBM Watson Knowledge Studio, Natural Language Classifier/Understanding (NLC/NLU)
IBM Watson Studio
IBM Object Storage
Dat
aset
(lag
u2
BC
L)M
od
elin
gIn
fere
nce
Cross Industry – Text Generation e.g. for Natural Language ProcessingWatson Studio
Quick Modelling• Epoch: 300• Loss: 0.2280• Acc: 91.83%
Generated Texts, given first few words
cintaku hanya untukmu dengan kesungguhankumimpi indahku yang kujalani sanubari ku
aku cinta dia tapi takdir memisahkanlupakanlah daku tinggalkan ini inginku mengulang
malam malam yang indah ada aku ada yangsayangku cintaku selalu
sedih hatiku bintangkamu kutunggu yang paling hot kuhujan rintik sang waktu bisa kitakuterima saja aku akan cinta yg
rembulan indah pun pernah muda saatnyaindahnya oh oh indahnya serasa
berbunga tinggalkan ini inginku mengulang
IBM Watson Studio
IBM Object Storage
Dat
aset
(Go
ogl
e N
ews)
Mo
del
ing
Infe
ren
ce
Cross Industry – Text Meaning in Context Watson Studio
Library• Gensim for NLP
+-
WomanKing
Man
Monarch, 61.89%
Princess, 59.02%
Queen, 71.18%*)
Crown Prince, 54.99%
Prince, 53.77%
Most Similar
Doesn’t Match
Tea
Blueberry, apple,
cucumber
Papaya, tea, guava
Orange, banana, tomato
*) Confidence level, utilizing pre-trained word2vec model based on GoogleNews dataset
H/W S/W (2.27)Socks Shoes (2.52)
Sky Earth (3.49)Saving Loan (3.88)
How Similar/Different
Watson Personality Insight
Source: SWG Insight, Edisi Q1 2018
Data Sourcetext, social media, etc
Human Capital - Identify Psychological TraitsWatson Personality Insight
Customerwho churned
Patterns:1) Very high usage of data and only data2) Usage of the phone only from 14-16th days3) Almost no usage during the 30 days period
1
2
3
One Customer
Day
s o
f th
e 3
0-d
ays
per
iod
10 Variables
# incoming calls
# outgoing calls
Sum duration incoming calls
Sum duration outgoing calls
Volume download
Volume upload
Duration data connection
# data connection
# MT SMS
# MO SMS
Max:1
Min: 0
1st day
2nd day
3rd day
Telco/Cross-Industry – Classification e.g. Customer’s Churn AnalyticsIBM Watson Studio or Watson Neural Network Modeler
Watson Visual RecognitionWatson Machine Learning
Data Source(video, images)
ClassificationAccording to
trained modelAnalyzes the photos that
customers submit a body damage claim, Applying the same
classification logic as highly experienced damage assessment
advisors
Identify selected part of a tree
Cross Industry - Tag, Classify, and Train Visual ContentIBM Watson Machine Learning, Watson Visual Recognition
Digit Prediction
Classificationdigit in range
[0..9]
Unseen Data
Neural Network modelcreated by IBM NN Modeler
Recognizing Handwritten Digit
Cross Industry - Classify Image/Text/AudioIBM Watson Studio or Watson Neural Network Modeler
Data prep made easy
Guided exploration
Understand outcomes
Share insights
Cross Industry - Advanced Analytics without complexityIBM Watson Analytics
Platform for Development, Deployment, & Model Management
Enabling Team Productivity & Collaboration: Domain Expert, Data Scientist, Data Engineer, Developer
End-to-end AI workflow: Connect/Access/ Search/Find/Prepare data for Analysis, Build & Train model, deploy model, monitor + analyze & manage
Available in Cloud, On-Premise, Desktop
Cross Industry - Advanced Analytics with FlexibilityIBM Watson Studio (Data Science Experience)
Financial ServicesDerive insights based on customer activity, investigate fraudulent behavior, create secure data transparency across global entities
Reduce readmissions, reduce time to comply with FDA regulations, reduce time to locate information for citizens and professionals
Increase document retrieval accuracy, improve search capabilities, share information while complying with sensitive data rules, dramatically increase agent productivity and customer satisfaction, reduce claim loss ratios and fraud, improve campaign effectiveness
Analyze social media and customer input for product development, identify product issues faster, reduce financial penalties and maintenance costs, improve supply chain consistency and deliver up to date information to procurement
Reduce time to find information for product investigations, improve access to patient and research data globally
Improve speed of regulation updates, connect citizens with public information, identify contraband activity, search and analyze complains, arrest records…
Improve product analysis, improve stocking decisions based on customer input, translate survey information in to actionable intelligence
Improve call center resolution rates, detect likely churn candidates, consolidate and normalize information after M&A activity
Healthcare
Insurance
Manufacturing
Pharma
Public Sector
Retail / CPG
Telco
Use Cases for Data Science – for Various Industries
If Your Data is Bad,Your Machine Learning tools are Useless
Data Scientists spend 80% of their time to cleanse data, before training the Predictive Model
It’s the Problem, data scientists complain about most
Harvard Business ReviewApril 2018
Data Science Journey
Harvard Business Review, April 2018
Determine Objective
Invest time for Quality
Data
Maintain Audit Trail
Assign PIC for Data Quality
Assign Independent
QA
Clarify Objectives of using Machine Learning, Assess whether you have the right data to support
Lowering the cost / Removing Bias / Improving existing decision process
Quality level, Data Cleansing, de-duplication
Possibly generating new data
About 6 months before building machine learning model
Original training data, data for training, steps to generate original data training data
Understand the bias & limitations
Set & enforce standard for data quality
Lead to find & eliminate root cause of error
Internal Quality Assurance (QA) department or 3rd
Party
Shop for data
Curate data
Manage policies
Shape data
Build dashboards
Build ML models
Auto-optimize models
Streaming pipelines
Build data apps
IBM Watson Data PlatformIntegrated, unified self-service experience
Andi Sama
Chief Information Officer
Formal EducationMaster: Bina Nusantara, 2006 - 2008
(Executive Program in Management)Master: STTI Benarif Indonesia, 1997 - 1998
(Computer Science)Bachelor: Bina Nusantara, 1987 - 1992
(Computer Engineering)
Executive Education2002: Owen Graduate School of Management
University of Vanderbilt, Nashville, TN USA(Banking Operation & Technology)
Professional Career2010-Now: Sinergi Wahana Gemilang
CIO- IBM Software VAD- Esri Partner- Autodesk Gold VAR
2000-2010: Mitra Integrasi KomputindoIT Integration Architect, CTO, Managed Services Head
1990-2000: IBM Indonesia, IBM Asia Pacific Developer, Technical Advisor, Technical Sales
Current Research InterestsData Science & Machine LearningInternet of ThingsBlockchainQuantum Computing
www.facebook.com/andisama www.facebook.com/SinergiWahanaGemilangwww.swgemilang.com SWG Insight App on Appstore & Playstore
Speaker Profile
More Info on…
IBM Data Science SWG Insight on Playstore or Appstore SWG Website or SWG Facebook page
Supervised Learning Unsupervised Learning Reinforcement Learninng
Semi Supervised Learning
Machine Learning Types
Source: Hands-On Machine Learning with Scikit-Learn and TensorFlow
Traditional Programming vs Machine Learning
Source: MIT Lecture 11 – Machine Learning, 6.0002, 2016
Traditional Programming Machine Learning
Computer Output
Data
Program/Algorithm
ComputerProgram/Algorithm
Data
Output
{[“Gunawan”, “M”, 45, 3200], [“May”, “F”, 31, 2400],[“Elle”, “F”, 27, 1925],[“Bob”, “M”, 36, 2375],…[“Boy”, “M”, 23, 1800]}
Y = ax + b
Y
Y
{[“Gunawan”, “M”, 45, 3200], [“May”, “F”, 31, 2400],[“Elle”, “F”, 27, 1925],[“Bob”, “M”, 36, 2375],…[“Boy”, “M”, 23, 1800]}
Y = ax + b
Machine Learning
TrainingEngine
(Train/Fit)
ModelTrainingData set
TestingData set
(Validation)
Target Output/Label
New (Unseen)Data set
Feedback/Iteration to minimize error (improve prediction accuracy)
70-80%
20-30%
Create Model Use Model
Inference Engine
(Predict)
AvailableData set
• “Target Classes” e.g. for Classification or “Target Values” for Regression if it is a Supervised Learning
• There are no target output in Unsupervised Learning (e.g. Clustering)
Output
Data Science Venn Diagram v2.0
Machine Learning as part of Data Science, is the
combination of computer science + math & statistics
disciplines.
Unicorn in the middle indicates that Data
Scientists are hard to find, as hard as unicorns.
ComputerScience
Math &Statistics
Subject Matter Expert
TraditionalSoftware
TraditionalResearch
MachineLearning
Unicorn(Data Scientist)
Data Science
©Stephen Geringer, Raleigh NC (2014)
ForwardPropagation
ForwardPropagation
“Human Face”
“Human Face”
Smaller, Varied N(New, Unseen Data)
= ?
Feedback/Iterationto minimize error
(improve prediction accuracy)
“Dog”Target Output Labels
USE (PREDICT)Deep Neural Network
(Architecture, Trained Weights)
Number ofepoch
GeneratesTrainedModel
Network Architecture (e.g. ResNet-152),
Initial Weights (e.g. Xavier),Activation Function (e.g. Sigmoid/ReLu)
BackwardPropagation
e.g. using SGD(Stochastic Gradient Descent)
TRAIN (FIT)Deep Neural Network
(Architecture, Weights,Activation Functions)
Learning rate,Update Weights
“Target Output Labels” e.g. for Classification or “Target Values” for Regression if within Supervised Learning
There are no target outputs in Unsupervised Learning (e.g. Clustering)
Large N(Training Dataset, Xi)
Typical Deep Learning Worklflow,with Supervising Learning (labeled data)