Stanford question & answer challenge
Ethical, legal, societal influences
Qualification problem
All preconditions?
Ramification problem
All effects of action?
Knowing that you do not know is the best.
Not knowing that you do not know is an illness.
- Laozi, 500-600 BCE
Fang, et al., 2015
Learn about abilities & failures
Performance
Successes & failures
p( fail | E, t)
Confidence
Image
H1
H2
H3
W1
W2
W3
W4
Input s
H3
Caption:
a man holding a tennis
racquet on a tennis court
H1
H2
H3
W1
W2
W3
Input t1
H3
W4
Deep learning about deep learning performance
Reliable predictions of performance: Known unknowns
Grappling with Open-World Complexity
Reliable predictions of performance: Known unknowns
Grappling with Open-World Complexity
Grappling with Open-World Complexity
Reliable predictions of performance: Known unknowns
Challenge of unknown unknowns
Expanded real-world testing
Algorithmic portfolios
Failsafe designs
People + machines
M
training data
M
real-world concepts
x= (𝑓1, … , 𝑓𝑘)
wrong label
high confidence
Conceptual incompleteness
cats
dogs
Lakkaraju, Kamar, Caruana, H, 2017.
Identifying classifier blindspots
M
training data
M
real-world concepts
x= (𝑓1, … , 𝑓𝑘)
wrong label
high confidence
cats
dogs
How to define & search regions of data space?
How to trade exploration and exploitation? Lakkaraju, Kamar, Caruana, H, 2017.
Identifying classifier blindspots
Conceptual incompleteness
M
training data
cats
dogs
M
training data
x= (𝑓1, … , 𝑓𝑘)
wrong label
high confidence
Partition
space by
attributes
White CatsWhite Dogs Brown DogsBrown Cats Lakkaraju, Kamar, Caruana, H, 2017.
Identifying classifier blindspots
Transfer learning
Learn from rich simulations
Learn generative models
Hospital A Hospital B
Hospital C
Site-specific data
Observations, definitions
Patients, prevalencies
Covariate dependencies
Transfer learning opportunity
Hospital C
Hospital B
J. Wiens, J. Guttag, H, 2015.
A: Community hosp: 10k pts/yr
B: Acute care & teaching: 15k/yr
C: Major teaching & research: 40k/yr
Site-specific data
Observations, definitions
Patients, prevalencies
Covariate dependencies
Transfer learning opportunity
Hospital A
Hospital C
Hospital B
J. Wiens, J. Guttag, H, 2015.
A: Community hosp: 10k pts/yr
B: Acute care & teaching: 15k/yr
C: Major teaching & research: 40k/yr
M. Gabel, R. Caruana, M. Philipose, O. Dekel
Less data with better features
ImageNet 1000, 1M photos
Cut off top layer
M. Gabel, R. Caruana, M. Philipose, O. Dekel
Less data with better features
ImageNet 1000, 1M photos
Cut off top layer
M. Gabel, R. Caruana, M. Philipose, O. Dekel
Less data with better features
ImageNet 1000, 1M photos
Cut off top layer
Raspberry Pi
Camera
Battery
Trillions of sessions in complex scenarios
Learn & evaluate core competencies
Learn to optimize action plans
Mapping
Planning
Next actions
Map
Plans
Stereo
algorithm
Depth Image
D. Dey, S. Sinha, S. Shah, A. Kapoor
CNN
Mapping
Planning
Next actions
Map
Plans
Stereo
algorithmCNN
Depth Image
D. Dey, S. Sinha, S. Shah, A. Kapoor
Learn expressive generative models
Generalize from minimal training sets
Harness physics
Multilevel variational autoencoder
Learn disentangled representations
Groups of observations latent models
Learning generative models
Vary IDVary style
D. Buchacourt, R. Tomioka, S. Nowozin, 2017
Smooth control over learned latent space
Inject physics to disentangle & generalize
Same?
Kulkarni, Whitney, Kohli & Tenenbaum, 2015
Inject physics to disentangle & generalize
Kulkarni, Whitney, Kohli & Tenenbaum, 2015
Illumination Nod Shake
Illumination Nod Shake
Inject physics to disentangle & generalize
Kulkarni, Whitney, Kohli & Tenenbaum, 2015
AI attack surfaces
Adversarial machine learning
Self-modification
Attacks on AI Systems
Goodfellow, et al.
Papernot, et al.
“Adverserial machine learning”
EnvironmentAction
Environment
AI system
e.g., see: Amodei, Olah, et al., 2016
State Perception
ReinforcementReward
Adversarial Attacks & Self-Modification
EnvironmentAction
Environment
AI system
e.g., see: Amodei, Olah, et al., 2016
State Perception
ReinforcementReward
Adversary
Adversarial Attacks & Self-Modification
EnvironmentAction
Environment
AI system
e.g., see: Amodei, Olah, et al., 2016
State Perception
ReinforcementReward
Adversary
Action
Adversarial Attacks & Self-Modification
EnvironmentAction
Environment
AI system
Amodei, Olah, et al., 2016H. 2016
State Perception
ReinforcementReward
Run-time verificationStatic analysis
Reflective analysisEnsure isolation * identify meddling * ensure operational faithfulness
Adversarial Attacks & Self-Modification
Models of people & tasks
Models of complementarity
Coordination of initiative
Models of people & tasks
Actions, services
E1 E2 E3
H1 H2
E4
Predictions about needs, goals
Models of world & people
Predictions about user beliefs
E2 E3
H1 H2
E4
E1 E2 E3
H1 H2
E4
Predictions about world
Actions
H. Barry, 1995
H. , Apacible, Sarin, Liao, 2005
H. Barry, 1995
Models of world & people
Complementarity
Complementarity
Complementarity
D. Wang, A. Khosla, R. Gargeya, H. Irshad, A.H. Beck, 2016
Identifying metastatic breast cancer
(Camelyon Grand Challenge 2016)
AI + Expert: 0.5%
85% reduction in errors.
Human is superior
Error: 3.4%
Complementarity
Machine
perception
Human
perception
Machine learning & inference
Kamar, Hacker, H., AAMAS 2012
ComplementarityLabel galaxies in Sloan Digital Sky Survey
(Galaxy Zoo)
~453 features
Machine learning & inference
Machine
perceptionHuman
perception
Kamar, Hacker, H., AAMAS 2012
ComplementarityLabel galaxies in Sloan Digital Sky Survey
(Galaxy Zoo)
~453 features
Machine learning & inference
Ideal fusion, stopping
Machine
perceptionHuman
perception
Kamar, Hacker, H., AAMAS 2012
Complementarity Full accuracy: 47% of human effort
95% accuracy: 23% of human effort
Designs for mix of initiatives
Machine learning & inference
Human
cognition
Machine
intelligence
C.E. Reiley, et al.
Initiative: Recognizing human goals, stateRecognizing intention
Padoy & Hager. ICRA 2011
van den Berg, et al, ICRA, 2010
Coordination of initiative
Trustworthiness and safety
Fairness, accuracy, transparency
Ethical and legal aspects of autonomy
Jobs and economy
Bernard Parker: rated high risk Dylan Fugett: rated low risk.
March 2017
A. Howard, C. Zhang, H., 2017
Machine learning “contact lens” for children
Science & engineering
Human-AI collaboration
AI, people, and society
Much to do