SLIDE 1
Stanford question & answer challenge Ethical, legal, societal - - PowerPoint PPT Presentation
Stanford question & answer challenge Ethical, legal, societal - - PowerPoint PPT Presentation
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
SLIDE 2
SLIDE 3
SLIDE 4
SLIDE 5
SLIDE 6
SLIDE 7
Stanford question & answer challenge
SLIDE 8
SLIDE 9
Ethical, legal, societal influences
SLIDE 10
Qualification problem
All preconditions?
Ramification problem
All effects of action?
SLIDE 11
SLIDE 12
Knowing that you do not know is the best. Not knowing that you do not know is an illness.
- Laozi, 500-600 BCE
SLIDE 13
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
SLIDE 14
Reliable predictions of performance: Known unknowns
Grappling with Open-World Complexity
SLIDE 15
Reliable predictions of performance: Known unknowns
Grappling with Open-World Complexity
SLIDE 16
SLIDE 17
SLIDE 18
Grappling with Open-World Complexity
Reliable predictions of performance: Known unknowns Challenge of unknown unknowns
SLIDE 19
Expanded real-world testing Algorithmic portfolios Failsafe designs People + machines
SLIDE 20
M
training data
M
real-world concepts x= (𝑔
1, … , 𝑔 𝑙)
wrong label high confidence
Conceptual incompleteness
cats dogs
Lakkaraju, Kamar, Caruana, H, 2017.
Identifying classifier blindspots
SLIDE 21
M
training data
M
real-world concepts x= (𝑔
1, … , 𝑔 𝑙)
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
SLIDE 22
M
training data cats dogs
M
training data x= (𝑔
1, … , 𝑔 𝑙)
wrong label high confidence Partition space by attributes White Cats White Dogs Brown Dogs Brown Cats
Lakkaraju, Kamar, Caruana, H, 2017.
Identifying classifier blindspots
SLIDE 23
Transfer learning Learn from rich simulations Learn generative models
SLIDE 24
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
SLIDE 25
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
SLIDE 26
- M. Gabel, R. Caruana, M. Philipose, O. Dekel
Less data with better features
ImageNet 1000, 1M photos Cut off top layer
SLIDE 27
- M. Gabel, R. Caruana, M. Philipose, O. Dekel
Less data with better features
ImageNet 1000, 1M photos Cut off top layer
SLIDE 28
- M. Gabel, R. Caruana, M. Philipose, O. Dekel
Less data with better features
ImageNet 1000, 1M photos Cut off top layer
SLIDE 29
Raspberry Pi Camera Battery
SLIDE 30
Trillions of sessions in complex scenarios Learn & evaluate core competencies Learn to optimize action plans
SLIDE 31
Mapping Planning Next actions Map Plans Stereo algorithm Depth Image
- D. Dey, S. Sinha, S. Shah, A. Kapoor
CNN
SLIDE 32
Mapping Planning Next actions Map Plans Stereo algorithm
CNN
Depth Image
- D. Dey, S. Sinha, S. Shah, A. Kapoor
SLIDE 33
Learn expressive generative models Generalize from minimal training sets Harness physics
SLIDE 34
Mu Multil tilevel evel variational iational autoencode toencoder
Learn rn di dise sent ntang ngled led repr present sentati ations
- ns
Groups ps of f obs bservations tions latent nt mo mode dels
Learning generative models
Vary ID Vary style
- D. Buchacourt, R. Tomioka, S. Nowozin, 2017
Smooth control over learned latent space
SLIDE 35
Inject physics to disentangle & generalize
Same? Kulkarni, Whitney, Kohli & Tenenbaum, 2015
SLIDE 36
Inject physics to disentangle & generalize
Kulkarni, Whitney, Kohli & Tenenbaum, 2015
Illumination Nod Shake
SLIDE 37
Illumination Nod Shake
Inject physics to disentangle & generalize
Kulkarni, Whitney, Kohli & Tenenbaum, 2015
SLIDE 38
AI attack surfaces Adversarial machine learning Self-modification
SLIDE 39
Attacks on AI Systems
Goodfellow, et al. Papernot, et al.
“Adverserial machine learning”
SLIDE 40
Environment
Action
Environment AI system
e.g., see: Amodei, Olah, et al., 2016
State Perception Reinforcement Reward
Adversarial Attacks & Self-Modification
SLIDE 41
Environment
Action
Environment AI system
e.g., see: Amodei, Olah, et al., 2016
State Perception Reinforcement Reward Adversary
Adversarial Attacks & Self-Modification
SLIDE 42
Environment
Action
Environment AI system
e.g., see: Amodei, Olah, et al., 2016
State Perception Reinforcement Reward Adversary Action
Adversarial Attacks & Self-Modification
SLIDE 43
Environment
Action
Environment AI system
Amodei, Olah, et al., 2016
- H. 2016
State Perception Reinforcement Reward
Run-time verification Static analysis
Reflective analysis
Ensure isolation * identify meddling * ensure operational faithfulness
Adversarial Attacks & Self-Modification
SLIDE 44
Models of people & tasks Models of complementarity Coordination of initiative
SLIDE 45
Models of people & tasks
Actions, services
E1 E2 E3 H1 H2 E4
Predictions about needs, goals
SLIDE 46
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
SLIDE 47
- H. Barry, 1995
Models of world & people
SLIDE 48
Complementarity
SLIDE 49
Complementarity
SLIDE 50
Complementarity
SLIDE 51
- 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
SLIDE 52
Machine perception Human perception Machine learning & inference
Kamar, Hacker, H., AAMAS 2012
Complementarity
Label galaxies in Sloan Digital Sky Survey
(Galaxy Zoo)
SLIDE 53
~453 features
Machine learning & inference Machine perception Human perception
Kamar, Hacker, H., AAMAS 2012
Complementarity
Label galaxies in Sloan Digital Sky Survey
(Galaxy Zoo)
SLIDE 54
~453 features
Machine learning & inference
Ideal fusion, stopping
Machine perception Human perception
Kamar, Hacker, H., AAMAS 2012
Complementarity
Full accuracy: 47% of human effort 95% accuracy: 23% of human effort
SLIDE 55
Designs for mix of initiatives Machine learning & inference
Human cognition Machine intelligence
SLIDE 56
C.E. Reiley, et al.
Initiative: Recognizing human goals, state Recognizing intention
SLIDE 57
Padoy & Hager. ICRA 2011 van den Berg, et al, ICRA, 2010
Coordination of initiative
SLIDE 58
Trustworthiness and safety Fairness, accuracy, transparency Ethical and legal aspects of autonomy Jobs and economy
SLIDE 59
SLIDE 60
Bernard Parker: rated high risk Dylan Fugett: rated low risk.
SLIDE 61
SLIDE 62
March 2017
- A. Howard, C. Zhang, H., 2017
Machine learning “contact lens” for children
SLIDE 63
SLIDE 64