Machine Learning
for the Automotive World
February 25th, 2020 Jeff Rehm
Topics
Behavioral Learning
The Vision
Challenges and Pitfalls
Implications for UX
Design
Predictions are hard –
particularly when they’re
about the future.
--Yogi Bera (?)
First, a Misconception
Having a vision is important-- there should always
be a goal to strive for.
But… we should take a sober, rational approach.
“If it’s built in Python, it’s Machine Learning.
If it’s built in PowerPoint, it’s Artificial Intelligence”
-- CloudMade Data Scientist
The idea of “Artificial Intelligence” in an automobile.
It is still a long, long way in the future.
Making predictions based on
the behavioral patterns and
preferences of an individual
Making predictions based on
large-scale aggregation of
behavior (speed, feature use,
etc.)
Making predictions based on
similarities to others
Cohort Learning
Personal LearningFleet Learning
Behavioral Learning
The Vision,
the Promise
JLR Video
https://www.youtube.com/watch?v=F923EuB06CI
Applications Today
Led by Luxury, but moving mainstream
JLR Smart Settings
Audi MIB2+
BMW Connected+
Hyundai Smart Cruise
Ford Mach-E
Mercedes MBUX
Typical Use Cases
Focused on the Journey
• Destination
• Route
• Departure time
• Arrival time
What they do on the way
• Media preferences
• Climate preferences
Built around the driver’s profile
Why has it taken so
long?
Connectivity
In-vehicle hardware
Backend systems at carmakers
Readiness of users
But mostly: Because it’s hard.
Configuration
Data Collection
Feature Extraction
ML
Data
Verification
Machine
Resource
Management
Analysis Tools
Process Management
Tools
Serving
Infrastructure
Monitoring
“Only a small fraction of real-world ML systems is composed of the ML code, as shown by the small
black box in the middle. The required surrounding infrastructure is vast and complex. “
Hidden Technical Debt in Machine Learning Systems
Google, Inc.
Why it’s so Hard
Challenges and Pitfalls
Going Rogue
Dangers: Feedback Loops, Non-
homogenous data, non-declared
consumers, etc.
à Technical debt. Potentially massive
amounts.
Privacy
• Trust – avoiding creepiness
• Data protection regulations
(GDPR, CCPA)
• Consent management
Disillusionment
• Expectations management
• Some people have predictable
behaviors. Others do not.
• UX Design is critical here
Coherency
ML must be managed across the entire
ecosystem (car, cloud, companion app)
• Consistency of predictions
• Holistic approach.
Big Challenge: The Business Case
What’s the ROI?
If there is a significant investment to
properly deploy ML, what’s the payback?
What is a prediction worth?
Valuing ML Solutions is not easy.
What did we learn?
• There are very few standalone “killer use-
cases”.
• High level “elegance” features are difficult
to value or monetize
• Instead it’s the accumulation of many
features that add up to provide synergy
Solve real problems
• Look further down Maslow’s Automaker’s
hierarchy of needs
à Focus on Safety, Emissions, Efficiency, Cost
of ownership, Retention, Costs of
development, manufacture, etc.
Deployment
Considerations
Architecture
• Where should the learning be done?
• Storage, memory, CPU resources, transmission,
cloud costs
• Where should the predictions be done?
• Latency? Context?
Managing data
• Validation
• Extraction
• Normalization across entire fleet
• Privacy
UX Design Paradigm Shift
Answer the ‘W’ questions:
• Where are they going? Who will go?
When will they leave? How will they
get there? What will they do on the
way?
Confidence? Predictions are
probabilities, not absolutes.
Integration: Non-deterministic
predictions must be integrated
into your deterministic UX logic.
This is where your skill enters in.
CONTEXT CONTEXT + INTENT
Example: 2 Variable Integration
Do Nothing
Confidence Level – Departure Time
ConfidenceLevel-Destination
100%0%
100%
Proactive push notification
Pre-populated suggestions
Proactive destination input
Future Applications
Adaptive UX
• Familiarity with area
• Cognitive workload
Feature on-boarding
• Cohort, Fleet, Personal
Preconditioning
• Climate
• EV batteries
Maintenance functions
• Performed at ideal time based on
predicted route, speed, etc.
Powertrain Efficiency
• Lean burn, ICE/BEV, Transmission
optimization
And on…
Take-Aways
Machine Learning is real.
Adds real value to the automaker and their customers.
Building a proper foundation is a must.
Maturity level is about to make a big jump.
It is truly a new frontier.
“To a man with a
hammer, everything
looks like a nail”
-- Mark Twain (?)
“AI is not magic pixie dust.
Predictions can provide value,
but there is no substitute for
good well-thought design.”
--Jeff
References
Yogi Berra quote
https://quoteinvestigator.com/2013/10/20/no-predict/
Land Rover's Self-Learning Intelligent Vehicle Video
https://www.youtube.com/watch?v=F923EuB06CI
Mercedes MBUX
https://www.mercedes-
benz.com/en/innovation/connected/mbux-mercedes-
benz-user-experience-revolution-in-the-cockpit/
Audi MIB2+ Infotainment
https://www.audi-mediacenter.com/en/the-new-audi-rs-
q8-the-most-sporty-q-12422/infotainment-and-audi-
connect-12432
Hyundai Smart Cruise (SCC-ML)
https://www.hyundainews.com/en-us/releases/2887
Ford Mustang Mach-E Launch
https://youtu.be/o0F9Uktpgtk?t=1309
Hidden Technical Debt in Machine Learning Systems
D. Sculley, Gary Holt, Daniel Golovin, et al , Google Inc.
https://papers.nips.cc/paper/5656-hidden-technical-debt-
in-machine-learning-systems.pdf
Mark Twain Quote
https://quoteinvestigator.com/2014/05/08/hammer-nail/

Machine Learning for the Automotive World

  • 1.
    Machine Learning for theAutomotive World February 25th, 2020 Jeff Rehm
  • 2.
    Topics Behavioral Learning The Vision Challengesand Pitfalls Implications for UX Design Predictions are hard – particularly when they’re about the future. --Yogi Bera (?)
  • 3.
    First, a Misconception Havinga vision is important-- there should always be a goal to strive for. But… we should take a sober, rational approach. “If it’s built in Python, it’s Machine Learning. If it’s built in PowerPoint, it’s Artificial Intelligence” -- CloudMade Data Scientist The idea of “Artificial Intelligence” in an automobile. It is still a long, long way in the future.
  • 5.
    Making predictions basedon the behavioral patterns and preferences of an individual Making predictions based on large-scale aggregation of behavior (speed, feature use, etc.) Making predictions based on similarities to others Cohort Learning Personal LearningFleet Learning Behavioral Learning
  • 6.
    The Vision, the Promise JLRVideo https://www.youtube.com/watch?v=F923EuB06CI
  • 7.
    Applications Today Led byLuxury, but moving mainstream JLR Smart Settings Audi MIB2+ BMW Connected+ Hyundai Smart Cruise Ford Mach-E Mercedes MBUX
  • 8.
    Typical Use Cases Focusedon the Journey • Destination • Route • Departure time • Arrival time What they do on the way • Media preferences • Climate preferences Built around the driver’s profile
  • 9.
    Why has ittaken so long? Connectivity In-vehicle hardware Backend systems at carmakers Readiness of users But mostly: Because it’s hard.
  • 10.
    Configuration Data Collection Feature Extraction ML Data Verification Machine Resource Management AnalysisTools Process Management Tools Serving Infrastructure Monitoring “Only a small fraction of real-world ML systems is composed of the ML code, as shown by the small black box in the middle. The required surrounding infrastructure is vast and complex. “ Hidden Technical Debt in Machine Learning Systems Google, Inc. Why it’s so Hard
  • 11.
    Challenges and Pitfalls GoingRogue Dangers: Feedback Loops, Non- homogenous data, non-declared consumers, etc. à Technical debt. Potentially massive amounts. Privacy • Trust – avoiding creepiness • Data protection regulations (GDPR, CCPA) • Consent management Disillusionment • Expectations management • Some people have predictable behaviors. Others do not. • UX Design is critical here Coherency ML must be managed across the entire ecosystem (car, cloud, companion app) • Consistency of predictions • Holistic approach.
  • 12.
    Big Challenge: TheBusiness Case What’s the ROI? If there is a significant investment to properly deploy ML, what’s the payback? What is a prediction worth? Valuing ML Solutions is not easy. What did we learn? • There are very few standalone “killer use- cases”. • High level “elegance” features are difficult to value or monetize • Instead it’s the accumulation of many features that add up to provide synergy Solve real problems • Look further down Maslow’s Automaker’s hierarchy of needs à Focus on Safety, Emissions, Efficiency, Cost of ownership, Retention, Costs of development, manufacture, etc.
  • 13.
    Deployment Considerations Architecture • Where shouldthe learning be done? • Storage, memory, CPU resources, transmission, cloud costs • Where should the predictions be done? • Latency? Context? Managing data • Validation • Extraction • Normalization across entire fleet • Privacy
  • 14.
    UX Design ParadigmShift Answer the ‘W’ questions: • Where are they going? Who will go? When will they leave? How will they get there? What will they do on the way? Confidence? Predictions are probabilities, not absolutes. Integration: Non-deterministic predictions must be integrated into your deterministic UX logic. This is where your skill enters in. CONTEXT CONTEXT + INTENT
  • 15.
    Example: 2 VariableIntegration Do Nothing Confidence Level – Departure Time ConfidenceLevel-Destination 100%0% 100% Proactive push notification Pre-populated suggestions Proactive destination input
  • 16.
    Future Applications Adaptive UX •Familiarity with area • Cognitive workload Feature on-boarding • Cohort, Fleet, Personal Preconditioning • Climate • EV batteries Maintenance functions • Performed at ideal time based on predicted route, speed, etc. Powertrain Efficiency • Lean burn, ICE/BEV, Transmission optimization And on…
  • 17.
    Take-Aways Machine Learning isreal. Adds real value to the automaker and their customers. Building a proper foundation is a must. Maturity level is about to make a big jump. It is truly a new frontier.
  • 18.
    “To a manwith a hammer, everything looks like a nail” -- Mark Twain (?) “AI is not magic pixie dust. Predictions can provide value, but there is no substitute for good well-thought design.” --Jeff
  • 19.
    References Yogi Berra quote https://quoteinvestigator.com/2013/10/20/no-predict/ LandRover's Self-Learning Intelligent Vehicle Video https://www.youtube.com/watch?v=F923EuB06CI Mercedes MBUX https://www.mercedes- benz.com/en/innovation/connected/mbux-mercedes- benz-user-experience-revolution-in-the-cockpit/ Audi MIB2+ Infotainment https://www.audi-mediacenter.com/en/the-new-audi-rs- q8-the-most-sporty-q-12422/infotainment-and-audi- connect-12432 Hyundai Smart Cruise (SCC-ML) https://www.hyundainews.com/en-us/releases/2887 Ford Mustang Mach-E Launch https://youtu.be/o0F9Uktpgtk?t=1309 Hidden Technical Debt in Machine Learning Systems D. Sculley, Gary Holt, Daniel Golovin, et al , Google Inc. https://papers.nips.cc/paper/5656-hidden-technical-debt- in-machine-learning-systems.pdf Mark Twain Quote https://quoteinvestigator.com/2014/05/08/hammer-nail/