From Zero to ML Hero for Underdogs - Amir Tabakovic
Download as PPTX, PDF1 like130 views
Amir Tabakovic presented at the Data Science Conference in Belgrade in November 2019, discussing the evolution of machine learning and the importance of integrating AI into businesses without relying solely on technology. He emphasized typical mistakes in AI transformation, such as cargo cult behavior and premature focus on infrastructure rather than practical applications. The presentation advocated starting small with pilot projects, building trained in-house teams, and developing a comprehensive AI strategy for successful implementation.
![AMIR TABAKOVIC
Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019
Think big
“There is nothing you can touch at Amazon, nothing on the website, nothing in the
business, that is not run by Machine Learning.
Everything you see […] everything is run by an underlying Machine Learning
system.”
General Manager Amazon ML](https://image.slidesharecdn.com/fromzerotomlheroforunderdogs-amirtabakovic-191226091313/85/From-Zero-to-ML-Hero-for-Underdogs-Amir-Tabakovic-12-320.jpg)
From Zero to ML Hero for Underdogs - Amir Tabakovic
- 2. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 From Zero to ML Hero for Underdogs Data Science Conference, November 2019 Amir Tabakovic @TABAKOVIC
- 3. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 About me Former Head of Market Development at PostFinance (Focus FinTech) VP of Business Development at BigML (Machine Learning Start-up) Independent strategy consultant, speaker & lecturer (University of Fribourg, Bern University of Applied Sciences, ESADE Barcelona, ICEMD Madrid) Escape Room designer: “learn about AI and master the challenge”
- 4. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Data driven products... Machine Learning Data Visualisation
- 5. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Expert Systems vs. Predictive Apps
- 6. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 "When I was a programmer, I was very good at figuring out all the algorithms and writing them all down. Today, I think I would try to figure out how to program a computer to learn something." Eric Schmidt Quote Source: https://www.straitstimes.com/tech/ma chine-learning-is-next-big-thing-in- programming Quote Pic: wikipedia.com
- 7. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Traditional Expert Systems: Expert + Programer ● Experts know how the decision / prediction is made (domain/business expertise) Start Stop Step 1 Step 2 Step 4 Step 5 Step 6B Step 3 Step 6A yes no ● Programmers know how to translate the expert’s knowledge into a computer program that makes the prediction (technical expertise) ● Constructing a system requires both (though they may be the same person)
- 8. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Predictive Applications: Data + Algorithm Illustration Source: BigML
- 9. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Classical Programming vs. Machine Learning Expert Systems Predictive Applications Rules Data Answer Classical Programming Answer Data Rules Machine Learning your priority nr. 1 in the next decade
- 10. How to start? … if you aren’t Amazon
- 11. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 (A bucket of) food for thought “You don’t make your company successful by buying a bucket of AI. It’s not about waiving an AI wand - there is no intrinsic value by itself. It’s about picking the right thing to do with AI.” Carl Hillier AI
- 12. AMIR TABAKOVIC Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Think big “There is nothing you can touch at Amazon, nothing on the website, nothing in the business, that is not run by Machine Learning. Everything you see […] everything is run by an underlying Machine Learning system.” General Manager Amazon ML
- 13. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Start small Unless your name is: , , , … stop acting like you are THE MASTER OF AI. Although there’s some experience in the market how to adopt best practices for big companies with huge budgets and access to extraordinary talent pools, there’s not much experience for SMEs when it comes to successful adoption of AI.
- 14. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 3 Typical mistakes in AI transformation Typical mistakes in AI transformation are: 1. cargo cult behaviour “Our data scientists are useless! All they do is sit around publishing papers.” 2. building data science teams with questionable expertise and fully disconnected from the rest of the organisation Super Ninja Jedi Data Scientist 3. focusing on the infrastructure projects dealing with pre-requirements for AI and delaying involvement with AI like projects until all issues are fixed “First we need a data lake”
- 15. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 1. Cargo cult
- 16. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Neural networks take over other ML methods
- 17. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Reinforcement learning is gaining momentum
- 18. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 2. Roles in AI/ML team when things go wrong... Super Ninja Jedi Data Scientist The Bad Boy Data Scientist The Baby Face Data Scientist The Interesting One Data Scientist The Rico Suave Data Scientist The Sporty One Data Scientist The “Meet the Parents” Data Scientist Icons made by itim2101 from flaticon.com Why so many data scientists are leaving their jobs https://www.kdnuggets.com/2018/04/why-data-scientists-leaving-jobs.html
- 19. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Roles in AI/ML team Source: https://hackernoon.com/top-10-roles-for-your-data-science-team-e7f05d90d961 Icons made by itim2101 from flaticon.com Data Engineer Analyst Applied ML Engineer Decision Maker “Data Scientist” Qualitative Expert Additional PersonnelData Architect Domain Expert Software Engineer Data Collection Specialist Data Product Manager UX Designer Graphic Designer Ethicist / Legal Expert ( )
- 20. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 3. “Infrastructure first” trap ● You will never be ready for Machine Learning (moving target) ● Data lake ≠ ML data lake (it’s helpful if you know in advance which features you will need for ML and generate them upfront) ● There will always be some additional data source to integrate ● It’s difficult to think about automation of ML workflows if there are no existing examples around ● ...
- 21. AMIR TABAKOVIC Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Life cycle AI/ML Tools (based on Wardley Mapping) Auto ML Data Science Tools Source: BigML
- 22. AMIR TABAKOVIC Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Sure, why not?
- 23. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Automation of AI/ML
- 24. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Advantages of Auto ML ● Increases productivity by automating repetitive tasks. This enables a data scientist to focus more on the problem rather than the models. ● Automating the ML pipeline also helps to avoid errors that might creep in manually. ● AutoML is a step towards democratizing machine learning by making the power of ML accessible to everybody. Source: https://heartbeat.fritz.ai/automl-the-next-wave-of-machine-learning-5494baac615f
- 25. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Alternatives: How to kickstart AI transformation Recommendation for AI transformation: 1. start small with a pilot project (to gain momentum) 2. build a well integrated in-house AI/ML-team – train the trainers 3. provide broad AI/ML training (for experts, for programers…) 4. develop an AI/ML strategy 5. communicate, iterate, deploy, communicate, iterate, deploy...
- 26. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Start your AI/ML journey with a smile NO-GO POSTPONABLE NO-BRAINER BRAINER ROI (impact / costs) FEASIBILITY (incdataavailability/deccomplexity)
- 27. Amir Tabakovic, experiens.ai, , Data Science Conference, Belgrade November 2019 Few examples for no-brainers & brainers No-brainers Brainers Predicting customer churn Chat-bot Predicting fraud (transactions, orders, warranty) Predicting success of early stage startups Predicting demand (propensity to buy, next best product, staffing) Stock-Trading-bot
- 28. We run out of yogurt Source::https://www.nusgram.com/media/BydPmTLiVL1
- 29. Before I run out of time
Editor's Notes
- #5: Abschluss der Vorstellung mit 2 Produkten, die ich selber initiiert habe.
- #9: Data replaces the expert Data is often more accurate than “expertise” Preparation reading: “Specialist Knowledge is Useless and Unhelpful” Using data requires (almost) no human expertise except how to collect the data The data may already be there Algorithms replace programmers Algorithms are faster than programmers, enabling iteration Algorithms are more modular than programmers Data offers the opportunity for performance measurement (which you would be doing in any case, right?) - whatever we hit we call target.
- #12: Your organisation could be seen as an aggregation of business rules: Production business rules Commercial business rules HR business rules ... Pre-digital concept - human intuition, best practices, common sense How frequently this rules are questioned? What’s is the alternative? Substitution of static human derived business rules with dynamic data driven rules
- #15: “Our data scientists are useless! All they do is sit around publishing papers.”
- #27: Success of the first project is more important than the absolute value (ROI) of the project.
- #29: Attentional Bias: This is the tendency to pay attention to some things while simultaneously ignoring others. When making a decision on which car to buy, you may pay attention to the look and feel of the exterior and interior, but ignore the safety record and gas mileage. Functional Fixedness: This is the tendency to see objects as only working in a particular way. If you don't have a hammer, you never consider that a big wrench can also be used to drive a nail into the wall. You may think you don't need thumbtacks because you have no corkboard on which to tack things, but not consider their other uses. This could extend to people's functions, such as not realizing a personal assistant has skills to be in a leadership role.