Pistoia Alliance Demystifying AI & ML part 2
1 like300 views
The document discusses a webinar series by the Pistoia Alliance focused on the introduction of AI and machine learning in life sciences. Key topics include investment trends in AI, data integrity, machine learning workflows, interpretability of models, and the application of conversational AI in business. It emphasizes the importance of regulatory considerations, ethics, and the need for skilled personnel in the rapidly evolving landscape of healthcare technology.
Pistoia Alliance Demystifying AI & ML part 2
- 1. 21 June, 2018 Demystifying AI – Part 2 An Introduction to AI in Life Sciences Pistoia Alliance Centre of Excellence for AI in Life Sciences and Health Prashant Natarajan (@BigDataCXO) Moderator: Nick Lynch
- 2. This webinar is being recorded
- 3. Poll Question 1: How much is your organisation planning to increase investment into AI/ML* in the next 2-3 years? (tools/platform, people etc) A. 0-25% B. >25% - 50% C. >50% - 75% D. >75% - 100% E. Not sure AI* = including machine learning/deep learning/chat bots)
- 4. ©PistoiaAlliance Webinars: AI in Life Sciences – Q2/Q3 2018 Pistoia Alliance Membership Introduction 4 • Webinar 1 (23 May 2018) Prashant Natarajan – A Brief History – Big Data/ML/DL/AI - fundamentals and concepts – Data Fidelity & NFR Framework – Best Practices from the Trenches – Q&A • Webinar 2: 21 June 2018 Prashant Natarajan – Big Data Analytics & AI - 2 sides of the same coin – A guided tour of learning algorithms for Healthcare – Real-life use cases in health & life sciences from the book Q & A – AI Solutions - Going Beyond Algorithms – Q & A • Webinar 3: July 2018 – (panel) – Real World Evidence, the Big Data Connection – The 3 P’s of RWE: Persons, Providers, and Pharma • Webinar 4 – State of the Art in AI with working examples • Etc – monthly Like to give a talk or panel? Boston Community Workshop Oct 2018
- 5. ©PistoiaAlliance Poll Question 2: Are you/is your organisation currently looking to hire additional AI/ML* experts or retrain existing staff? A. Yes, now or soon B. Yes in the next 12 months C. Yes but later than 12 months D. No E. Don’t Know AI* = including machine learning/deep learning/chat bots etc)
- 6. Prashant Natarajan • Senior Director of AI Applications at H2O.ai, Mountain View, CA, USA (www.h2o.ai) • Undergraduate degree in Chemical Engineering; Master’s in Technical Communications & Linguistics; PhD courses in Logic & Cognitive Psychology; AT&T- Yahoo Chancellor’s Fellow • 18+ years in health sciences industry – providers, pharma, payers, patients • H2O.ai; Oracle Health Sciences; McKesson; Healthways; Siemens • Lead author or contributor to books on big data analytics, business intelligence, cancer, machine learning, AI (best-sellers in 2012, 2017, 2018) • Co-Faculty Instructor, Stanford University School of Medicine, Palo Alto, CA • Industry Advisor, CA Initiative to Advance Precision Medicine/San Francisco VA @BigDataCXO | prashant.natarajan@gmail.com | www.BigDataCXO.com
- 7. ©PistoiaAlliance Agenda 721 June, 2018 • Considerations for Life Sciences • ML 102 • TIE – Interpretability & Explainability • Conversational AI: Bot Basics • Q & A
- 8. ©PistoiaAlliance Consideration for Life Sciences 821 June, 2018 • Regulations and policy • Innovation in a regulated environment • TIE it up • Organization and structural challenges in Life Sciences • Resourcing • Data fidelity and labeling • MDM is critical as is data governance • Ethics and privacy – human and machine morality are not the same. Does a machine have morals? • Clear demarcation or sharing of human & machine- learning/CIA responsibilities when failure happens
- 9. Machine Learning 102 Mastering the Basics Sources: www.H2O.ai Driverless AI overview “Demystifying Big Data and Machine Learning for Healthcare” (Taylor & Francis, 2017), Natarajan et al. “Principles of Data Wrangling” (O’Reilly, 2017), Rattenbury et al. AWS Sagemaker Developer Guide Prashant Natarajan
- 10. ©PistoiaAlliance Typical Enterprise Machine Learning Workflow ModelModel Building Features Target Modeling Table Data Quality & Transformation Data Integration + Driverless AI Copyright 2018 H2O.ai Inc. All rights
- 11. ©PistoiaAlliance ML Workflows: from Data to Deployment
- 12. ©PistoiaAlliance Data Preparation & Wrangling 1221 June, 2018 • Ingest Data from RDBMS, files, distributed DBs, etc – describe data - assess data utility • Create & manage metadata • Profile data – grain, structure, data fidelity, temporality, scope • Pre-visualization and outlier analysis • Refine data – mastering, structuring (changing form or schema), enriching (adding new info via joins, unions, derived data), transforming (cleansing, addressing missing/invalid values) • Create production data for training and use/build automated ML systems to process all the way to the scoring pipeline (or) visualization
- 13. ©PistoiaAlliance Training & Scoring in H2O’s Driverless AI "Confidential and property of H2O.ai. All rights reserved" Data Processing Model Tuning Feature Engineering Final Model Training Scoring Pipeline
- 14. ©PistoiaAlliance Deployment & Tracking 1421 June, 2018 • Monitor Ongoing Performance - How will you monitor the performance of your algorithm on an ongoing basis? Data drifts and systems evolve. • Look for ability to connect to your existing visualization – verify interpretability – make it easy for data scientists/IT/business to collaborate via results and code • Keep Track Of Your Model Changes - Always track the revision of your model and report it with your results. As you improve different parts of your data analytics pipeline, you will want to go back and re-analyze data. Recording which model was used at which time helps you understand what to recalculate.
- 15. ©PistoiaAlliance ML Workflows: from Data to Deployment
- 17. ©PistoiaAlliance Interpretability *Source: https://christophm.github.io/interpretable-ml-book/interpretability-importance.html 1721 June, 2018 TIE: Interpretability is the degree to which a human can understand the cause of a decision (Miller 2017)* • If the ML model performs well, can’t we just trust it? • “The problem is a single metric, such as classification accuracy, is an incomplete description of most real- world tasks” (Doshi-Velez & Kim) • What v why/how of predictions: knowing the “why” can help you understand more about the problem, data, biases, leaks, debug/audit, and why a model might fail • Facilitate learning and satisfy human curiosity • The model becomes the source of insights and knowledge – not just the raw data. Hence, interpretability becomes important • Interpretability is not the same as explainability
- 18. ©PistoiaAlliance Interpretability Source: https://christophm.github.io/interpretable-ml-book/interpretability-importance.html 1821 June, 2018 If the ML model is interpretable and explainable, we can check for the following traits: • Fairness: why was “x” denied a credit limit upgrade? Is there a racial bias in the data? • Privacy: ensuring sensitive data in the information is tracked and protected • Robustness: testing that small changes in inputs don’t lead to big changes in prediction • Trust: humans trust a system that explains decisions compared to a black box When don’t we need interpretability? • Problem is too well-studied • Model has no significant impact • Enable “gaming” of the ML system
- 19. ©PistoiaAlliance Conversational AI Examining Bot Basics Sources: Demystifying Big Data & Machine Learning for Healthcare (Natarajan et al, CRC Press, 2017) “Designing Bots: Creating Conversational Experiences” (Amir Shervat, O’Reilly Press 2017) Prashant Natarajan
- 20. ©PistoiaAlliance AI & Bots: the Connections 2021 June, 2018 Conversational AI & Bots • Most bots are powered by ML/AI – though not all of them • Designing a great conversation is orthogonal, in most cases, to the decision to use AI or another technology • What can AI do for bots today? – Natural Language Understanding (extracting & converting free text to entities) – Conversation mgmt. and context switching – Computer vision and image recognition – Prediction – finding patterns and predicting outcomes based on past data – Sentiment analysis – understanding emotional state • Bot types: personal v team, super v domain-specific, business v consumer, text v voice, Net New Service v New Interfaces
- 21. ©PistoiaAlliance Anatomy of a Bot 2121 June, 2018 • Bot anatomies are important given that the primary purpose of a bot is to recognize and help accomplish human intent • Anatomical features of a bot include – Branding, personality and human involvement – AI – Conversation management: onboarding, flows, feedback/error handling, help and support – Rich interactions via files, audio, images, buttons, helpful links, emojis, typing indicators, Web views – Context and memory – Engagement methods: notifications, user-led, subscriptions
- 22. ©PistoiaAlliance Some Use Cases 2221 June, 2018 • Bot anatomies are important given that the primary purpose of a bot is to recognize and help accomplish human intent • Conversational commerce – FB, Alexa, etc • Bots for business – Slackbot, GitHub ChatOps • Productivity and coaching – Lark, AHA, etc • Alerts and notifications • Router between humans (Uber, Lyft, scheduling bots) • Customer service and FAQs • 3rd party integration bots (Slack and CRM) • Games and entertainment • Brand bots
- 23. ©PistoiaAlliance Poll Question 3: How important do you feel FAIR* data principles are to ensuring successful outputs from AI projects ? A. Very important B. Important C. Neutral D. Not important E. Not Very important FAIR : Findable Accessible, Interoperable & Reusable
- 24. ©PistoiaAlliance Audience Q&A Please use the Question function in GoToWebinar
- 25. ©PistoiaAlliance RWD and AI – how can work they together? The next Pistoia Alliance CoE AI Webinar: Date: TBD July 2018 check http://www.pistoiaalliance.org/events/ for the latest information
- 26. info@pistoiaalliance.org @pistoiaalliance www.pistoiaalliance.org Thanks for your engagement