https://www.youtube.com/watch?v=HAfLCTRuh7U&t=19s
Eduard MIRESCU
Feb. 2018
ARTIFICIAL INTELLIGENCE
AND THE FUTURE OF QA
ARTIFICIAL INTELLIGENCE (1950) – branch of computer science concerned
with making computers behave like humans.
MACHINE LEARNING (1959) – area of AI that is focused on developing
principles and techniques for automating the acquisition of knowledge ||
learn without being explicitly programmed.
PROJECT ALPHAGO
AlphaGo is an AI that plays Go. It is trained using deep neural networks and reinforced learning. It was shown
hundreds of thousands of Go matches and the AI learned everything there is know about Go.
2015 – AlphaGo defeated the European Go Champion 5-0.
2016 – AlphaGo defeated the Worldwide Go Champion 4-1.
PROJECT ALPHAGO ZERO
AlphaGo Zero is the next evolution of AlphaGo. Instead of learning from existing matches, the AI is given the
rules of the game and an empty board. It plays against itself millions of times and learns from its own
mistakes. In a matter of days, the AI discovered strategies that took the Go community thousands of years to
discover.
2017 – AlphaGo Zero beats AlphaGo 100-0.
SUPPORT/CALL
CENTER CHATBOT
VOICE RECOGNITION CHATBOTS
SIRI CORTANA ALEXA
SCIENTIFIC CHATBOTGENERALIST CHATBOT
https://www.youtube.com/watch?v=seCUU5FlWIQ
https://www.youtube.com/watch?v=jPajqAJWiNA
HEALTHCARE
- Analyze Tests Results – X-Rays, CT Scans, bloodwork
- Treatment Design – improves a patient treatment’s plan
- Virtual Nurses – chatbots that can provide customized
information for patient’s family
- Drug Creation – identify the closest medication that can be
easily modified to cure or prevent a disease
- Precision Medicine – predict cancer or cardiovascular
diseases using body scans and DNA analysis
- Health Monitoring – predict strokes
IMAGE RECOGNITION / SURVEILLANCE IMAGE RECOGNITION / SELF DRIVING CARS
https://www.youtube.com/watch?v=aE1kA0Jy0Xg https://www.youtube.com/watch?v=VG68SKoG7vE&
PREDICTIVE ANALYTICS & CLASSIFICATION
Stocks and cryptocurrencies value
Sport events result
Corruption predictions
Risks identification
Sales forecasting
House value prediction
Fraud prevention
Toxic comments and fake news analysis
Iceberg identifications and path prediction
Gifts matching
Entertainment Recommendations
Traffic predictions and optimization
product team
users
system
input data
output data
define system rules
feedback
system
product team
users
output data
input data
AI MODEL
DATASET
select data
configure model
define system rules
feedback
TRADITIONAL SYSTEM AI BASED SYSTEM
DEMO
SENSITIVITY TO
FEEDBACK
BIASES
UNEXPECTED
RESULTS
BIASES
In machine learning, a bias usually occurs when the AI model is trained using an overfitted or underfitted set of data.
The root cause may be within the data itself or how the data was prepared:
- excluded/under-represented segments
- incorrect labels
- imbalanced data preparation
- irrelevant data present
ALL EXISTING DATA DATA WE HAVE DATA WE VALIDATE WITH
REAL WORLD BIAS EXAMPLES
- US courtrooms use a re-offending assessment risk
algorithm that is biased against black people
- New Zealand’s passport system incorrectly interprets
Asian photos as having their eyes closed
- SAS’s performance evaluation algorithm got more
than 200 teachers fired as it used insufficient data
- YouTube’s speech-to-text system does not recognize
women’s voices
- Google Photos automatically tagged some pictures
containing black people as “gorillas”
- Flicker tagged Dachau concentration camp as “sport”
or “jungle gym”
UNEXPECTED RESULTS
“We can build these models, but we don’t know how they work.”
- AI algorithms are hard to understand (black box)
- Current design techniques are built around logic, where we know the expected result
- AI systems are probabilistic, as opposed to deterministic
Traditional eCommerce site:
If a product costs 10 USD and I add 2 to my shopping cart I will have to pay 20 USD.
eCommerce site that uses AI to modify prices to in real time to offer maximize sells and profits:
If a product usually costs 10 USD, the AI might sell it to you for 9 USD, and to someone else for 11 USD, but if you add 3 in the
shopping cart you might have to pay 28 USD and someone else 30 USD.
UNEXPECTED RESULTS
prepare the data:
- it must include all features and
classes as per model design
- all business segments are
properly represented
- have at least two sets (one for
test and validation, and another
for the actual tests)
- test data set should include
anomalies from both technical
and business perspectives
train/validate the model:
- verify if results are within the
expected boundaries
test the model:
- feed it unseen/new data
- analyze scoring
- create result visualizations to
understand biases
- verify the results of anomalous
data
- feed it more data an recheck
initial results
SENSITIVITY TO FEEDBACK
BLACK BOX ATTACKS USING ADVERSARIAL EXAMPLES
AIs are vulnerable to adversarial examples: malicious inputs modified to yield erroneous
model outputs, while appearing unmodified to a human observer.
SELF DRIVING CARS
Level Capabilities
0 – No Automation driver has to do everything
1 – Drive Assistance either automated steering
OR
acceleration/deceleration
2 – Partial Automation both steering AND
acceleration/deceleration
3 – Conditional Automation can drive itself, but driver
needs to interfere in some
cases depending on the
environment
4 – High Automation can drive itself even if the
driver does not
appropriately respond to
requests to interfere
5 – Full Automation can drive itself in any
conditions that a human
driver could
AI TESTING
Level Capabilities
0 – No AI tester has to do everything
1 – Automation repetitive tasks can be
scripted and executed
automatically
2 – Partial AI AI learns from tester’s data
and improves it
3 – Conditional AI test cases are generated
and executed
automatically, but the
tester needs to interfere
4 – High AI test cases are generated
and executed
automatically, tester needs
to validate results
5 – Full AI test cases are generated
and executed
automatically
LOCALIZATION TESTING
diacritice.ai – identifies letters that should be diacritics and automatically replaces them with the correct ones
UI TESTING
applitools - uses image recognition to identify humanely visible anomalies in layout
BUG TRIAGE
IBM DeepTriage (Research) – AI model capable of identifying the most appropriate developer to fix a defect
PREDICTIVE ANALYTICS / PROCESS IMPROVEMENT
- identify similar/duplicate and unique test cases
- predict defect clusters and automatically prioritize
automated test case execution on that areas
- impact assessment for future user stories
- identify risks that can appear in next iterations
- automatic traceability between requirements, test
cases and defects
https://www.youtube.com/watch?v=DnwSbLZjgdc
self healing automated scriptsself generating test casesautomated exploratory testing
CONCEPTS
2030
+14%
16T USD
26% - 7T
15% - 4T
11% - 3T
10% - 1T
Impact 0-3 3-7 7+
3.7
3.7
3.3
3.2
3.1
3.0
2.2
2.2
37%
35%
41%
41%
47%
54%
39%
14%
23%
47%
59%
41%
36%
38%
44%
83%
40%
18%
0%
18%
17%
8%
17%
3%
6.5M 12M 25M
2017 2018 2020
Thank you!
Q&A
AI Bias Real World Examples:
- http://vamboozled.com/breaking-news-victory-in-court-in-houston/
- https://www.technologyreview.com/s/607955/inspecting-algorithms-for-bias/
- https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
- https://www.reuters.com/article/us-newzealand-passport-error/new-zealand-passport-robot-tells-applicant-of-asian-
descent-to-open-eyes-idUSKBN13W0RL
- http://edition.cnn.com/2009/TECH/12/22/hp.webcams/index.html
- https://www.pcworld.com/article/209708/Is_Microsoft_Kinect_Racist.html
- http://www.businessinsider.com/google-tags-black-people-as-gorillas-2015-7
- https://www.theguardian.com/technology/2015/may/20/flickr-complaints-offensive-auto-tagging-photos
- https://mashable.com/2017/11/30/google-translate-sexism/#0ZWvyZCrZsqG
- https://mashable.com/2017/10/25/google-machine-learning-bias/?utm_cid=a-seealso#y5rC05vWfmq6
- https://www.technologyreview.com/s/604087/the-dark-secret-at-the-heart-of-ai/
Unexpected Behaviors:
- http://www.testingtrapezemagazine.com/wp-content/uploads/2017/04/TestingTrapeze-2017-April.pdf
- https://www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist
- https://arxiv.org/pdf/1312.6199.pdf
- https://blog.openai.com/adversarial-example-research/
- https://arxiv.org/pdf/1412.6572.pdf
- https://arxiv.org/pdf/1602.02697.pdf
Financials and Predictions:
- https://www.gartner.com/smarterwithgartner/top-trends-in-the-gartner-hype-cycle-for-emerging-technologies-2017/
- https://www.statista.com/statistics/607716/worldwide-artificial-intelligence-market-revenues/
- https://evansdata.com/reports/viewRelease.php?reportID=9
- https://www.pwc.com/gx/en/issues/analytics/assets/pwc-ai-analysis-sizing-the-prize-report.pdf
- http://www3.weforum.org/docs/FOP_Readiness_Report_2018.pdf
Applications of AI in QA:
- http://esanu.name/vitalie/?p=861
- https://diacritice.ai/
- https://applitools.com/
- https://www.infosys.com/IT-services/validation-solutions/service-offerings/Pages/machine-learning-qa.aspx
- https://www.stickyminds.com/interview/future-software-testing-ai-interview-jason-arbon
- http://www.tothenew.com/blog/artificial-intelligence-in-software-testing/
- https://en.wikipedia.org/wiki/Autonomous_car
- http://bugtriage.mybluemix.net/#chrome

Artificial Intelligence and QA

  • 1.
  • 2.
    Eduard MIRESCU Feb. 2018 ARTIFICIALINTELLIGENCE AND THE FUTURE OF QA
  • 3.
    ARTIFICIAL INTELLIGENCE (1950)– branch of computer science concerned with making computers behave like humans. MACHINE LEARNING (1959) – area of AI that is focused on developing principles and techniques for automating the acquisition of knowledge || learn without being explicitly programmed.
  • 4.
    PROJECT ALPHAGO AlphaGo isan AI that plays Go. It is trained using deep neural networks and reinforced learning. It was shown hundreds of thousands of Go matches and the AI learned everything there is know about Go. 2015 – AlphaGo defeated the European Go Champion 5-0. 2016 – AlphaGo defeated the Worldwide Go Champion 4-1. PROJECT ALPHAGO ZERO AlphaGo Zero is the next evolution of AlphaGo. Instead of learning from existing matches, the AI is given the rules of the game and an empty board. It plays against itself millions of times and learns from its own mistakes. In a matter of days, the AI discovered strategies that took the Go community thousands of years to discover. 2017 – AlphaGo Zero beats AlphaGo 100-0.
  • 5.
    SUPPORT/CALL CENTER CHATBOT VOICE RECOGNITIONCHATBOTS SIRI CORTANA ALEXA SCIENTIFIC CHATBOTGENERALIST CHATBOT https://www.youtube.com/watch?v=seCUU5FlWIQ
  • 6.
  • 7.
    HEALTHCARE - Analyze TestsResults – X-Rays, CT Scans, bloodwork - Treatment Design – improves a patient treatment’s plan - Virtual Nurses – chatbots that can provide customized information for patient’s family - Drug Creation – identify the closest medication that can be easily modified to cure or prevent a disease - Precision Medicine – predict cancer or cardiovascular diseases using body scans and DNA analysis - Health Monitoring – predict strokes
  • 8.
    IMAGE RECOGNITION /SURVEILLANCE IMAGE RECOGNITION / SELF DRIVING CARS https://www.youtube.com/watch?v=aE1kA0Jy0Xg https://www.youtube.com/watch?v=VG68SKoG7vE&
  • 9.
    PREDICTIVE ANALYTICS &CLASSIFICATION Stocks and cryptocurrencies value Sport events result Corruption predictions Risks identification Sales forecasting House value prediction Fraud prevention Toxic comments and fake news analysis Iceberg identifications and path prediction Gifts matching Entertainment Recommendations Traffic predictions and optimization
  • 10.
    product team users system input data outputdata define system rules feedback system product team users output data input data AI MODEL DATASET select data configure model define system rules feedback TRADITIONAL SYSTEM AI BASED SYSTEM
  • 11.
  • 12.
  • 13.
    BIASES In machine learning,a bias usually occurs when the AI model is trained using an overfitted or underfitted set of data. The root cause may be within the data itself or how the data was prepared: - excluded/under-represented segments - incorrect labels - imbalanced data preparation - irrelevant data present ALL EXISTING DATA DATA WE HAVE DATA WE VALIDATE WITH
  • 14.
    REAL WORLD BIASEXAMPLES - US courtrooms use a re-offending assessment risk algorithm that is biased against black people - New Zealand’s passport system incorrectly interprets Asian photos as having their eyes closed - SAS’s performance evaluation algorithm got more than 200 teachers fired as it used insufficient data - YouTube’s speech-to-text system does not recognize women’s voices - Google Photos automatically tagged some pictures containing black people as “gorillas” - Flicker tagged Dachau concentration camp as “sport” or “jungle gym”
  • 15.
    UNEXPECTED RESULTS “We canbuild these models, but we don’t know how they work.” - AI algorithms are hard to understand (black box) - Current design techniques are built around logic, where we know the expected result - AI systems are probabilistic, as opposed to deterministic Traditional eCommerce site: If a product costs 10 USD and I add 2 to my shopping cart I will have to pay 20 USD. eCommerce site that uses AI to modify prices to in real time to offer maximize sells and profits: If a product usually costs 10 USD, the AI might sell it to you for 9 USD, and to someone else for 11 USD, but if you add 3 in the shopping cart you might have to pay 28 USD and someone else 30 USD.
  • 16.
    UNEXPECTED RESULTS prepare thedata: - it must include all features and classes as per model design - all business segments are properly represented - have at least two sets (one for test and validation, and another for the actual tests) - test data set should include anomalies from both technical and business perspectives train/validate the model: - verify if results are within the expected boundaries test the model: - feed it unseen/new data - analyze scoring - create result visualizations to understand biases - verify the results of anomalous data - feed it more data an recheck initial results
  • 17.
  • 18.
    BLACK BOX ATTACKSUSING ADVERSARIAL EXAMPLES AIs are vulnerable to adversarial examples: malicious inputs modified to yield erroneous model outputs, while appearing unmodified to a human observer.
  • 19.
    SELF DRIVING CARS LevelCapabilities 0 – No Automation driver has to do everything 1 – Drive Assistance either automated steering OR acceleration/deceleration 2 – Partial Automation both steering AND acceleration/deceleration 3 – Conditional Automation can drive itself, but driver needs to interfere in some cases depending on the environment 4 – High Automation can drive itself even if the driver does not appropriately respond to requests to interfere 5 – Full Automation can drive itself in any conditions that a human driver could AI TESTING Level Capabilities 0 – No AI tester has to do everything 1 – Automation repetitive tasks can be scripted and executed automatically 2 – Partial AI AI learns from tester’s data and improves it 3 – Conditional AI test cases are generated and executed automatically, but the tester needs to interfere 4 – High AI test cases are generated and executed automatically, tester needs to validate results 5 – Full AI test cases are generated and executed automatically
  • 20.
    LOCALIZATION TESTING diacritice.ai –identifies letters that should be diacritics and automatically replaces them with the correct ones
  • 21.
    UI TESTING applitools -uses image recognition to identify humanely visible anomalies in layout
  • 22.
    BUG TRIAGE IBM DeepTriage(Research) – AI model capable of identifying the most appropriate developer to fix a defect
  • 23.
    PREDICTIVE ANALYTICS /PROCESS IMPROVEMENT - identify similar/duplicate and unique test cases - predict defect clusters and automatically prioritize automated test case execution on that areas - impact assessment for future user stories - identify risks that can appear in next iterations - automatic traceability between requirements, test cases and defects https://www.youtube.com/watch?v=DnwSbLZjgdc
  • 24.
    self healing automatedscriptsself generating test casesautomated exploratory testing CONCEPTS
  • 27.
    2030 +14% 16T USD 26% -7T 15% - 4T 11% - 3T 10% - 1T Impact 0-3 3-7 7+ 3.7 3.7 3.3 3.2 3.1 3.0 2.2 2.2 37% 35% 41% 41% 47% 54% 39% 14% 23% 47% 59% 41% 36% 38% 44% 83% 40% 18% 0% 18% 17% 8% 17% 3% 6.5M 12M 25M 2017 2018 2020
  • 28.
  • 30.
    AI Bias RealWorld Examples: - http://vamboozled.com/breaking-news-victory-in-court-in-houston/ - https://www.technologyreview.com/s/607955/inspecting-algorithms-for-bias/ - https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing - https://www.reuters.com/article/us-newzealand-passport-error/new-zealand-passport-robot-tells-applicant-of-asian- descent-to-open-eyes-idUSKBN13W0RL - http://edition.cnn.com/2009/TECH/12/22/hp.webcams/index.html - https://www.pcworld.com/article/209708/Is_Microsoft_Kinect_Racist.html - http://www.businessinsider.com/google-tags-black-people-as-gorillas-2015-7 - https://www.theguardian.com/technology/2015/may/20/flickr-complaints-offensive-auto-tagging-photos - https://mashable.com/2017/11/30/google-translate-sexism/#0ZWvyZCrZsqG - https://mashable.com/2017/10/25/google-machine-learning-bias/?utm_cid=a-seealso#y5rC05vWfmq6 - https://www.technologyreview.com/s/604087/the-dark-secret-at-the-heart-of-ai/ Unexpected Behaviors: - http://www.testingtrapezemagazine.com/wp-content/uploads/2017/04/TestingTrapeze-2017-April.pdf - https://www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist - https://arxiv.org/pdf/1312.6199.pdf - https://blog.openai.com/adversarial-example-research/ - https://arxiv.org/pdf/1412.6572.pdf - https://arxiv.org/pdf/1602.02697.pdf
  • 31.
    Financials and Predictions: -https://www.gartner.com/smarterwithgartner/top-trends-in-the-gartner-hype-cycle-for-emerging-technologies-2017/ - https://www.statista.com/statistics/607716/worldwide-artificial-intelligence-market-revenues/ - https://evansdata.com/reports/viewRelease.php?reportID=9 - https://www.pwc.com/gx/en/issues/analytics/assets/pwc-ai-analysis-sizing-the-prize-report.pdf - http://www3.weforum.org/docs/FOP_Readiness_Report_2018.pdf Applications of AI in QA: - http://esanu.name/vitalie/?p=861 - https://diacritice.ai/ - https://applitools.com/ - https://www.infosys.com/IT-services/validation-solutions/service-offerings/Pages/machine-learning-qa.aspx - https://www.stickyminds.com/interview/future-software-testing-ai-interview-jason-arbon - http://www.tothenew.com/blog/artificial-intelligence-in-software-testing/ - https://en.wikipedia.org/wiki/Autonomous_car - http://bugtriage.mybluemix.net/#chrome

Editor's Notes

  • #3 - prezentare
  • #4 Joc de dame Joc de X&0 IBM DeepBlue / Kasparov
  • #6 Personal Assistants Trecere catre anti-scamming
  • #8 - Radiologia va disparea
  • #9 - Camere foto cu Face Recognition la inceputul anilor 2000
  • #10 FinTech foloseste inca din anii 90 AI pentru predictia fluctatilor la bursa, mai nou si pentru crypto Betting industry foloseste AI pentru predictia rezultatelor meciurilor Magazin de jucarii din SUA Corruption in Spain
  • #11 - Exemplu Booking A/B Testing pentru colectarea de feedback
  • #24 Predictive analytics Exemplu identificarea riscurilor folosind date din JIRA
  • #26 Explicatie schema Unde e Romania
  • #27 Explicatie schema Unde suntem acum Exemplu Apple
  • #28 PIB-ul Mondial Impact in industrii Fintech primii care vor ajunge la 100% Numarul de dezvoltatori