Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

AI Symposium Keynote Manila, 2017

133 views

Published on

Technical Introduction to AI and Data with context of education and business models.

Published in: Data & Analytics
  • Be the first to comment

  • Be the first to like this

AI Symposium Keynote Manila, 2017

  1. 1. Ikhlaq Sidhu, content author Ikhlaq Sidhu Founding Faculty Director Sutardja Center for Entrepreneurship & Technology Department of Industrial Engineering & Operations Research IEOR Emerging Area Professor Award Innovation & Entrepreneurship and AI
  2. 2. Ikhlaq Sidhu, content author What is happening at Sutardja Center (SCET)?
  3. 3. Ikhlaq Sidhu, content author Answer #1: Lots of students and activity Undergraduate: Over 1500 Graduate: Over 100 Executive: Over 100 Global Partners – Now 10
  4. 4. Ikhlaq Sidhu, content author Answer# 2: Many News Stories about our Curriculum 5/27/2017 Meat substitutes are on the curriculum at UC Berkeley - San Francisco Chronicle http://www.sfchronicle.com/business/article/Meat-substitutes-are-on-the-curriculum-at-UC-10881462.php 1/5 By Jonathan Kauffman | January 24, 2017 | Updated: January 25, 2017 2:44pm 0 Meat substitutes are on the curriculum at UC Berkeley Biz & Tech IMAGE 1 OF 4 Ricardo San Martin, a chemical engineering professor, leads the Challenge Lab researching meat substitutes. Photo: Scott Strazzante, The Chronicle 5/27/2017 Meat substitutes are on the curriculum at UC Berkeley - San Francisco Ch Banking Un Course - Stu Singapore Most UC Berkeley students will tell you that they’re shooting for an A. But the 45 young men and women packing a Barrows Hall classroom this Monday were pursuing more ambitious goals: saving the world, and perhaps winning $5,000 in the process. The students are enrolled in a four-credit Challenge Lab at the Sutardja Center for Entrepreneurship & Technology, a practicum that pits teams against one another to develop the most innovative plant-based meat. At the lab’s second gathering, there were no Tofurky samples on hand. Instead, Christie Lagally, a senior scientist with the Go the students a PowerPoint crash course on the reasons the world needs m “Factory farming allows us to have an affluence of meat,” she told them, several dozen charts illustrating the downside of our omnivorous appetit figures, animal agriculture produces up to 24 percent of greenhouse gase About 7 Stories about our plant based meat focus area including Vice Magazine and SF Chronicle
  5. 5. Ikhlaq Sidhu, content author Answer #3: And a lot more.. Actually, we have lost track..
  6. 6. Ikhlaq Sidhu, content author Our Approach
  7. 7. Ikhlaq Sidhu, content author Misconception: We used to think that learning business and management would help technology innovators Reading business cases studies on innovation Studying business frameworks Management Practices and financial statements Waiting for a great idea Making a business plan Making a presentation to raise funds 6 things are not the main ingredient to deploy innovation or start ventures!
  8. 8. Ikhlaq Sidhu, content author Our Model Has Adapted: Business training is not the key. The New formula is: > depth in an valued area > entrepreneurial “behaviors and mindset”” Our programs and projects provide this. Innovation Behaviors and Mindset “Psychology of Innovation” Skill in a Core Area Too Narrow Street Smart, but lacking depth High Potential
  9. 9. Ikhlaq Sidhu, content author • Wide Comfort Zone • Generate Trust • Good Connectors • Inductive Learning: Experiments and Reflection • Self awareness and Emotional Intelligence • And a few more .. Innovation Behaviors and Mindset Skill in a Core Area Too Narrow Street Smart, but lacking depth High Potential What are the Behaviors and Mindsets? Taught in situation, during the journey
  10. 10. Ikhlaq Sidhu, content author My newest course: IEOR 135 Applied Data Science (Data-X)
  11. 11. Ikhlaq Sidhu, content author Sample Project from the First Data-X Course • Detection of fake news • Prediction of long-term energy prices to solve am Wall Street problem • Prediction applications stock market, sports betting, and more • AI for Crime detection, traffic guidance, medical diagnostics, .. etc • A version of Zillow that is recalculated with the effects of AirBnB income • and many more…
  12. 12. Ikhlaq Sidhu, content author Propose Low Tech Solution (1) Brainstorm Challenge and Validate (4) Demo or Die (1) Execute * Iterate BMoE Reflections Agile Sprint (8) Insightful Story Solution How the Data-X Course Works: Team: typically 5 students, with available advisor network
  13. 13. Ikhlaq Sidhu, content author The Data-X System View: It’s more than ML Web Scrape Possible Input Code Blocks Download Crawl … Stream or Poll Social Net / IoT Application with Automated Decisions Algorithm Options w/ Tables/Matrix Prediction / Classification Test, train, split Keep state Pandas: Short Term Storage Long Term Storage: SQL and File Formats (JSON, CSV, Excel) Web Possible Output Code Blocks Email Control Decision … Chatbot Feedback from External System (World) Pre- process Natural Languag e, State Features Blockchain (public ledger or crypto-lock)APIs, Services APIs, Services ML
  14. 14. Ikhlaq Sidhu, content author Our Newest Course contributed to IEOR’s core Area: IEOR 135 Applied Data Science (Data-X) 5/27/2017 Data-X: An Experimental Course Model that is Working - UC Berkeley Sutardja Center Search 8 MAY 2017Data-X: An Experimental Course Model that is Working by Ikhlaq | posted in: ariti cial intelligence, big data, Sutardja Center News, undergraduate classes | I’d like to start by congratulating all the teams that participated in our rst Data-X course this spring. We just watched the nal presentations, and it has been a great experience. Three months ago, we were just introducing the basic frameworks. And now, by the end of the semester, the projects have included running code and insightful approaches to topics such as: Detection of fake news Prediction of long-term energy prices to solve a Wall Street problem Prediction applications for the stock market and sports betting AI for Crime detection, traf c guidance, and medical diagnostics A version of Zillow that is recalculated with the effects of AirBnB income and many more… Students presenting at Data-X nals   These are technically dif cult projects, not to mention creative and inspiring. Everyone has come up a very large learning curve. I want to thank Kevin Bozhe Li and Alexander Fred Ojala for being part of the teaching team. And our guests, such as Rob von Behren from Google who spoke on TensorFlow and entrepreneurs like Antonio Vitti who brought real life problems and context to the course. Today, the world is literally reinventing itself with Data and AI. However, neither leading companies nor the world’s top students have the complete knowledge set or access to the full networks they need to participate in this newly developing world. Data-X is a UC Berkeley course and a global project designed to x this problem. Undergraduate Courses and Certi cate Graduate Program The Berkeley Method of Entrepreneurship BMoE Bootcamp Engineering Leadership Professional Program Startup Semester at Berkeley Innovation Collider 2017 Spring Newton Lecture Series About the Center Login Recent Posts Free Ventures Demo Day: From Seed to Startup Why You Should Learn Data-X Engineered In uence: Weak Data, Machine Learning & Behavioral Economics Students serve up next generation plant-based seafood Data-X: An Experimental Course Model that is Working Home About Courses People Insight News Explore Contact Q: What Are you getting from this class? A: I feel like I'm really learning how powerful data science tools can be. When we were brainstorming project ideas, I didn't think any of the ideas were feasible. However, with each week, I'm learning how pre existing libraries and tools can be easily used and combined to create really powerful products.
  15. 15. Ikhlaq Sidhu, content author We are developing a large-scale, holistic, data-related skill base • The Data-X Project is program and open course model • Offers deep skills, the powerful open source CS tools, and the real-life applications • Ready to scale and include more stakeholders
  16. 16. Ikhlaq Sidhu, content author Data, AI, and Business Models
  17. 17. Ikhlaq Sidhu, content author Scoring Wine Wine quality = 12.145 + 0.00117 x (winter rainfall )+ 0.0614 x (ave growing season temperature) – 0.00386 x (harvest rainfall), Oren Ashenfelter, Princeton. Now used by Christies Auction House
  18. 18. Ikhlaq Sidhu, content author Real-life Example: ZestCash • “All data is credit data” Online Loan Application Name: JOE SMITH Online Loan Application Name: Joe Smith The data says: greater credit risk! The data says: lesser credit risk! Reference: Shomit Ghose Example: Data and information is a competitive advantage
  19. 19. Ikhlaq Sidhu, content author Harrah’s Casino: Knowing your customer Service provider of Gambling and Casinos Entry Card Pain points Intervention Reference: Supercrunchers Example
  20. 20. Ikhlaq Sidhu, content author 1. Knowing your customer, better targeting and relationship. E.g. Target, Disney, Netflix 2. Improving physical product or servicer with complimentary information: E.g. UPS, FedEx 3. Data-driven reliability or security E.g. GE, BMW, Siemens 4. Information Brokers, Arbitrage, and Trading Opportunities: E.g. Investment funds. 5. Improving the customer journey/experience.. E.g. Harrah’s 6. Functional Applications: HR/Hiring, Operations etc.. Eg Walmart, Baseball, Sports 7. Efficiency or better performance per dollar cost. E.G. General IT, SAP, etc 8. Risk Management, regulation, and compliance Eg. Compliance 360 Top 8 Business Models Using Data
  21. 21. Ikhlaq Sidhu, content author Top Business Models for Using Data 1. Knowing your customer, leading to better targeting and relationship. E.g. Target, Disney 2. Information based better services. E.g. UPS, FedEx 3. Data driven reliability. E.g. GE and Siemens 4. Information Brokers, Arbitrage, and Trading Opportunities: Investment funds. 5. Improving the customer journey/experience.. E.g. Harrahs 6. Functional Applications: HR/Hiring, Operations etc.. 7. Efficiency or Better Performance per dollar cost 8. Risk Management, regulation, and compliance Usage Models • Efficiency (save money) • Wallet Share (top customers spend more time and money with you) • Brand alignment (It reinforces how people think positively about the company) Value to Business Customers More Value
  22. 22. Ikhlaq Sidhu, content author The two key components of a business are resources (assets) and information (data) = + Less value over time More value Over time Information and automated decisions If you buy data, then everyone else has it also.
  23. 23. Ikhlaq Sidhu, content author University Researcher Perspective
  24. 24. Ikhlaq Sidhu, content author Misconception Work in Lab for 5 years Show World They Love it And adopt it
  25. 25. Ikhlaq Sidhu, content author Misconception Work in Lab for 5 years Show World They Love it And adopt it Instead: Invite World (Industry) to collaborate with you Let them tell you where the industry will be in 5 years Intersect with their roadmap in 2-3 years They Love it And adopt it a) And train yourself and your students to have the corresponding mindsets and behaviors during the journey b) work on deployment first and effectiveness first
  26. 26. Ikhlaq Sidhu, content author Data and AI Fundamentals
  27. 27. Ikhlaq Sidhu, content author Basic Concept of Big Data * Data Wrangling * In Production
  28. 28. Ikhlaq Sidhu, content author Human Interpretation of Data Human Machines Large Sets of Data Insight
  29. 29. Ikhlaq Sidhu, content author An ML High Level Framework Objects Events/Experi ments People/Custo mers Products Stocks … In Real Life Features, but also loss of information In Sample Out of Sample Person 1 Person 2 Person 3 . . . Person N Characteristics Patterns Models Predictions Similarities Differences Distance Some data has observed results
  30. 30. Ikhlaq Sidhu, content author An ML High Level Framework Objects Events/Experi ments People/Custo mers Products Stocks … In Real Life Features, but also loss of information In Sample Out of Sample Person 1 Person 2 Person 3 . . . Person N Characteristics Patterns Models Predictions Similarities Differences Distance CS: Table Math: Matrix X, which is N rows – each person m columns, each feature (age, salary, ..) X = Some data has observed results
  31. 31. Ikhlaq Sidhu, content author A Fundamental Idea: From Table to Score Element F1 F2 F3 A 4 2 2 B 4.5 1.5 3 C 3 3 5 D 1 2 2 E 3 1.5 5 F 3.5 3.5 1 .. .. .. .. F(X) Element Credit Score A 552 B 381 C 760 D 330 E 452 F 678 .. .. X Y
  32. 32. Ikhlaq Sidhu, content author A Fundamental Idea: From Table to N- Dimensional Space A B CD E F G H 1 2 3 4 5 5 4 3 2 1 Element F1 F2 F3 A 4 2 2 B 4.5 1.5 3 C 3 3 5 D 1 2 2 E 3 1.5 5 F 3.5 3.5 1 .. .. .. .. X =
  33. 33. Ikhlaq Sidhu, content author Clustering to Classification 33 A B CD E F G H Feature 1 Feature 2 1 2 3 4 5 5 4 3 2 1 Target customers? Pictures of Cats and Dogs Speech recognition Recognize Letters: A, B, C..
  34. 34. Ikhlaq Sidhu, content author Machine Learning: Learning from Data Input Data = Matrix X Customer 1: [Name, income, x, y, .. Features ..z] Customer 2: [Name, income, x, y, .. Features ..z] Customer N: [Name, income, x, y, .. Features ..z] Output Data = Column Vector Y Customer 1: [20] Customer 2: [60] Customer N: [05] Purchases/year, repaid loan, … Target: What is F(X) = Y a formula that we don’t know Sample data (training): (x1,y1) (x2,y2) … (xm,ym) we have this Algorithm A from H H: Hypothesis Set: All possible algorithms or formulas Find G(x) which is approx. F(x) a) Supervised ML – as shown b) Unsupervised - no training data c) Reinforced learning – done by simulation
  35. 35. Ikhlaq Sidhu, content author KNN / K Means Illustration 12/19/2016 How to choose machine learning algorithms | Microsoft Docs https://docs.microsoft.com/en-us/azure/machine-learning/machine-learning-algorithm-choice 17/19 A data set is grouped into 5 clusters using K-means There is also an ensemble one-v-all multiclass classifier, which breaks the N-class classification problem into N-1 two-class classification problems. The accuracy, training time, and linearity properties are determined by the two-class classifiers used. + Options https://docs.microsoft.com/en-us/azure/machine-learning/machine-learning-algorithm-choice Tip Flavors of machine learning Supervised Letter recognition. To download and print a diagram that gives an overview of the capabilities Studio, see Overview diagram of Azure Machine Learning Studio capabilitie Supervised learning algorithms make predictions based on a set of e historical stock prices can be used to hazard guesses at future prices training is labeled with the value of interest—in this case the stock p learning algorithm looks for patterns in those value labels. It can use might be relevant—the day of the week, the season, the company's f industry, the presence of disruptive geopolicitical events—and each Illustration Source: KNN Method: Find the k nearest images and have them vote on the label (i.e. take the mode)
  36. 36. Ikhlaq Sidhu, content author K Means / KNN Illustration 12/19/2016 How to choose machine learning algorithms | Microsoft Docs https://docs.microsoft.com/en-us/azure/machine-learning/machine-learning-algorithm-choice 17/19 A data set is grouped into 5 clusters using K-means There is also an ensemble one-v-all multiclass classifier, which breaks the N-class classification problem into N-1 two-class classification problems. The accuracy, training time, and linearity properties are determined by the two-class classifiers used. + Options https://docs.microsoft.com/en-us/azure/machine-learning/machine-learning-algorithm-choice Tip Flavors of machine learning Supervised Letter recognition. To download and print a diagram that gives an overview of the capabilities Studio, see Overview diagram of Azure Machine Learning Studio capabilitie Supervised learning algorithms make predictions based on a set of e historical stock prices can be used to hazard guesses at future prices training is labeled with the value of interest—in this case the stock p learning algorithm looks for patterns in those value labels. It can use might be relevant—the day of the week, the season, the company's f industry, the presence of disruptive geopolicitical events—and each Illustration Source: KNN Method: Find the k nearest images and have them vote on the label (i.e. take the mode) from sklearn.neighbors import KNeighborsClassifier knn = KNeighborsClassifier(n_neighbors = 3) knn.fit(X_train, Y_train) Y_pred = knn.predict(X_test) acc_knn = round(knn.score(X_train, Y_train) * 100, 2) acc_knn # or compare Y_pred with Y_test
  37. 37. Ikhlaq Sidhu, content author Our experiment with the Titanic Data Set Model Score Random Forest 86.76 Decision Tree 86.76 KNN 84.74 Support Vector Machines 83.84 Logistic Regression 80.36 Linear SVC 79.01 Perceptron 78.00 Naive Bayes 72.28 Stochastic Gradient Decent 72.28 More Accuracy Generally more training time More risk of overfitting Less Accuracy Generally less computation
  38. 38. Ikhlaq Sidhu, content author Accuracy Increases with amount of Training Data
  39. 39. Ikhlaq Sidhu, content author X Y X Y Input Data = Matrix X Customer 1: [Name, income, x, y, .. Features ..z] Customer 2: [Name, income, x, y, .. Features ..z] Customer N: [Name, income, x, y, .. Features ..z] Output Data = Column Vector Y Customer 1: [20] Customer 2: [60] Customer N: [05] Purchases/year, repaid loan, … Target: What is F(X) = Y a formula that we don’t know
  40. 40. Ikhlaq Sidhu, content author Neural net results are close t human results
  41. 41. Ikhlaq Sidhu, content author This means accuracy Trade-offs: Training complexity/time vs Accuracy Sometimes good enough is good enough
  42. 42. Ikhlaq Sidhu, content author All our course materials: • Slides • Code samples • References are available at data-x.blog Free to use. Course Material
  43. 43. Ikhlaq Sidhu, content author Anticipating the Next Industrial Revolution Industrial Revolution 1.0 Industrial Revolution 2.0 • Winner was whoever made something most cheaply • Leveraged scale • Winner will be whoever makes best sense of the data • Leveraging scale Shomit Ghose
  44. 44. Ikhlaq Sidhu, content author Data and AI Effects Everything We Know Every Business: Will deconstructed by Data, AI, and Automated Decisions Society: Danger of even larger gap between the highly skilled and the lessor skilled Government: Must adapt to a new level of transparency and efficiency or face trouble from their people People: Will change their behaviors (like cell phone to the power 10). Work like balance, social structure, and hybrid human machine.
  45. 45. Ikhlaq Sidhu, content author Contact: Ikhlaq Sidhu Chief Scientist & Founding Director, Center for Entrepreneurship & Technology Faculty Director, Engineering Leadership Professional Program (ELPP) IEOR Emerging Area Professor, UC Berkeley sidhu@berkeley.edu scet.berkeley.edu

×