Machine Learning 101 (Or, Why You Should Care About ML) Cory Janssen, Co-Founder, AltaML
1 “I have a hard time thinking of an industry that I don’t think AI will transform in the next several years”
- Andrew Ng, Founder Deeplearning.Ai
2 What is ML? A Few Definitions
Artificial Intelligence Machine Learning Deep Learning
Replicating part, all, or more Enabling computers to learn without Building multiple layers of than human intelligence being explicitly programmed (usually abstraction on datasets to with large datasets) construct higher-level meaning
3 What is ML? Not Traditional
Data Traditional Answers Programming Rules
Data Machine Rules Learning Answers
4 Data Network More Users Effect
Better Care More Data
Smarter Algorithms
5 Moore’s Law
Vacuum Electromechanical Relay Transistor Integrated Circuit Tube 1010
• Computer power/$ 108 increases exponentially, with a doubling time of 18- 24 months (5 years = 10x) 106 • This trend has held for at least sixty years and will 4 10 continue for the Why foreseeable future 102 Now?
1 It’s About Calculations Calculations per Second per$1,000
10-2 The Hardware
10-4
10-6 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 2020 Year 6 Edmonton Edge
The AI community in Edmonton may have gotten off to a quiet start, but it is now viewed as a world leader in the field. – Financial Post
7 1992
World’s First Officially Sanctioned Man-Machine Computer Championship
Jonathan Schaeffer (Dean of Science at UofA)
8 1997
Deep Blue Defeats Kasparov
Murray Campbell, UofA grad played key role
9 2007
Checker is Solved
Jonathan Schaeffer in Edmonton
10 2015
Heads-up limit Texas Hold’em Solved
Deepstack Team, led by Dr. Michael Bowling
11 2016
AlphaGo Defeats Lee Sedol
David Silver and Aja Huang (pictured) attended UofA
12 2017
Google’s Deepmind Opens in Edmonton
Demis Hassabis, Deepmind Co-founder, with UofA
13 Examples - Deepmind
Cut energy used for cooling by 40% at Google’s data centers
14 Examples – Outcome Prediction
Produced a model with 93%-94% accuracy in predicting in-hospital mortality versus 86% for traditional methods
15 Examples – Harnessing Omics Data
Screened 72 million compounds to select potential anti-cancer drugs
16 Examples – Intelligent CDSS
Identifies and triages suspected stroke patients from CT images, in 95% of cases alerting a specialist an average of 52 minutes earlier
17 Examples – AI Enabled Imaging
Eliminates manual tasks by a radiologist to render an MRI scan in 9x less time
18 Examples – Diabetic Retinopathy
87.2% sensitivity for detecting diabetic retinopathy within 20 seconds of image capture
19 The Difficulty…
Unless we make a small bet, the project will fail because of culture and buy-in.
20 ML Project Elements
Data
Machine Learning Team
Domain Expertise
21 How We Work
Roadmap and Model Development Integration Deep Dive and MVP
Analyze potential use-cases and the Focus on delivering an initial ML solution to Focus on integrating ML solution into the viability of an ML solution. selected business problems. organization.
1) Problem 1) Data Pipeline Development 1) Model Sustainability 2) Metrics 2) Exploratory Data Analysis 2) APIs 3) ROI 3) Model Development 3) Front-end Solution 4) Data 4) Minimum Viable Product 4) Integrated Solution 5) Model
Goal: Provide detailed, Goal: Develop initial model/MVP to Goal: Integrate a functioning ML prioritized ML roadmap illustrate effectiveness of ML solution solution
3 Months 3-6 Months 3 Months
22 "It’s the first inning. It might even be the first guys up at bat. We're on the edge of the golden age [of AI]."
- Jeff Bezos
23 Cory Janssen 780-982-5472 [email protected] 24