Machines are now bettering people at a range of specific intuition tasks. How are they doing this and what might be next?
What does this mean for the development and governance of intelligent products and services that we access as consumers and citizens, and what are the implications for the organisations that provide them?
David explores the impact across the product lifecycle, from customer perception, to job design, technology development, knowledge management, risk and ethics.
14. WHAT IS AN ARTIFICIAL NEURAL NETWORK?
12
Weights of
connections
between nodes
repeatedly
adjusted as
network “learns”
training set of
labelled images
1.
TRAINING
Network can
classify novel
images with
“learned” labels
2.
USAGE
22. EXPLOSION OF MICRO-INTUITORS
20
There is almost nothing we
can think of that cannot
be made new, different, or
more valuable by infusing
it with some extra IQ.
The business plans of the
next 10,000 start-ups are
easy to forecast:
take X and add AI.
Kevin Kelly
The Inevitable
topbots.com
133 Enterprise AI Startups
23. HUMANS AND MACHINES COMPLEMENTING EACH OTHER
21
MACHINES
Wider learners
Scalable thinkers
HUMANS
Faster learners
Flexible thinkers
24. DISTRIBUTED MACHINE INTUITION
22
Apple iOS Core ML
Google Federated Learning
CLOUD TRAINING DEVICE INTUITION
AGGREGATED DATA PRIVATE DATA
FEDERATING LEARNING
29. UNDERSTAND CUSTOMER PERCEPTION
27
Just 7 percent of respondents would
trust a robot with their savings,
versus the 14 percent willing to submit
to a machine for heart surgery.
Andy Maguire
COO HSBC Europe
31. TAKE A HUMAN-CENTRED APPROACH
29
GOOGLE BRAIN
People + AI Research
institute
Prototype with Real People Instead of an Algorithm
Machine Learning Doesn’t Solve Everything
Design with the System’s Failure in Mind
Get Feedback, Forever
33. BUSINESS VS GAMEPLAY
31
BUSINESS GAMEPLAY
Goal
Environment
Limited Resources
Moves
Taking Turns
Scoring
Results
More
freedom.
Act ethically!
Less
freedom.
Follow the
rules!
Necessarily
complex
Simplification
reveals
insight
36. THE NEW NEW PRODUCT DEVELOPMENT GAME
34
OBJECTIVES
PRODUCT
Validate
RULES +
SOFTWARE
ARCHITECTURE
Research &
Analyse
Develop
(PEOPLE)
CODE
Verify
Deploy
37. THE NEW NEW NEW PRODUCT DEVELOPMENT GAME
35
OBJECTIVES
PRODUCT
Validate
RULES +
SOFTWARE
ARCHITECTURE
Research &
Analyse
DATA SET +
NETWORK
ARCHITECTURE
Curate
Data
Deploy
Develop
(PEOPLE)
CODE
Verify
Deploy
Train
(MACHINES)
MODEL
Verify
38. DATA CURATION
36
DATA SET +
NETWORK
ARCHITECTURE
Curate
Data
OBJECTIVESYour existing data
• Needs 10,000-100,000 samples
• Normalisation
• Bias treatment
Generating new data
• Secondary app
• Simulation
• Generative networks
39. MACHINES TRAINING
37
DATA SET +
NETWORK
ARCHITECTURE
Train
(MACHINES)
MODEL
Verify
Concerns
• Number crunching
• Fine-tuning & debugging
• Evolving network architecture
Approaches
• Dedicated hardware
• Pre-trained models
• Adversarial networks
• High-level frameworks & automation
40. SKILLS SETS
38
Diverse teams
• More robust approaches
• Less risk of inadvertent bias
Data scientists
• High-level direction - hybrid approaches
• Deep technical expertise
• Risk and quality assurance
Architecture & Development
• AI API
• Integrate data pipelines & UI
• Develop automation tools
• Run network iterations and training cycles
TOOLS & FRAMEWORKS (EG)
Platforms
• TensorFlow - general purpose and
cross-architecture
• Caffe - specifically vision - includes
Model Zoo with pre-trained models
High-level frameworks
• KERAS - python based flexible
network definition
• DeepLearning.scala - differentiable
functional programming
• PyTorch - evolving and flexible
51. MACHINES BETTERING HUMANS, NOT JUST BESTING HUMANS
45
As Fan’s losses piled up
against AlphaGo, [he] came
to see Go in an entirely new
way. Against other humans,
he started winning more.
Cade Metz
Wired
Just as machines made human
muscles a thousand times
stronger, machines will make
the human brain a thousand
times more powerful.
Sebastian Thrun
Google X
54. CHANGING JOB DESIGN
47
Alexandra Heath
Head of Economic Analysis Department, RBA
Carlos Perez
Intuition Machine, University of Massachusetts Lowell
Jobs that use automation as a tool
Jobs that use humans as a safety valve against
automation failure
Jobs that interpret the decisions of machines
Jobs that design human-machine interfaces
Jobs that design automation to manipulate
human behaviour
58. HUMAN FAILURE MODES
50
Training Set Bias
Google image search: “faces in things”@thisismoonlight
Execution Variability
You are anywhere between two and six times as likely to
be released if you're one of the first three prisoners
considered versus the last three prisoners considered.
https://www.theguardian.com/law/2011/apr/11/judges-lenient-break
Fooled by Hidden Heuristics
59. NO EXPLANATION - THE “SEMANTIC GAP”
51
BLACK BOX
INTUITION
KNOW WHY
KNOW HOWSCENARIO
60. SOCIETAL IMPLICATIONS
52
According to recent reports, every
Chinese citizen will receive a so-
called ”Citizen Score” [based in
part on deep learning against
Baidu history], which will
determine under what conditions
they may get loans, jobs, or travel
visa to other countries.
https://www.scientificamerican.com/article/will-democracy-survive-big-data-and-artificial-intelligence/
61. SOME RESPONSES TO RISK & ETHICAL CHALLENGES
53
WEAPONS OF MATH
DESTRUCTION
1. Are your objectives
aligned with your
customer’s?
2. Is your model’s operation
opaque?
3. Is it “scaled” from a
similar application, or
likely to “scale” in turn to
related applications?
4. Does your model create
its own reality with
feedback loops?
EU RIGHT TO
EXPLANATION
General Data Protection
Regulation to take effect
2018. The law will effectively
create a “right to
explanation” for users
about whom algorithmic
decisions were made.
In its current form, the GDPR’s
requirements could require a
complete overhaul of
standard and widely used
algorithmic techniques.
SELF-DRIVING CARS
Mercedes will save
occupants as a priority -
they have taken a pre-
meditated position.
Volvo will accept liability for
any incident involving one
of their vehicles in
autonomous mode.
63. OUTPUT LAYER
55
Machines are outperforming humans in narrow intuition tasks.
This is due to a confluence of recent developments,
and innovation continues at incredible pace.
This delivers enormous improvement potential, and has significant implications
for how we design and develop products, and how we build and manage organisations.
There are potentially huge benefits for society,
but ethics and risk to be managed.
Understanding machine intuition better will help us better manage these developments,
and ultimately help us understand humans better.