In this talk, Bernease Herman speaks about recent explainable ML research
Presented on 06/06/2019
**These slides are from a talk given at Rsqrd AI. Learn more at rsqrdai.org**
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
Rsqrd AI: Recent Advances in Explainable Machine Learning Research
1. Recent advances in
explainable machine
learning research
Bernease Herman, Research Scientist
University of Washington eScience Institute and
Paul G. Allen School of Computer Science & Engineering
June 6, 2019 - Allen Institute for Artificial Intelligence
2. Recent advances in
explainable machine
learning research
Bernease Herman, Research Scientist
University of Washington eScience Institute and
Paul G. Allen School of Computer Science & Engineering
June 6, 2019 - Allen Institute for Artificial Intelligence
Speed run edition
9. Defining model explainability
No single definition within community.
“component of interpretable modeling process informing
how model works in understandable form” - Me
“process of giving explanations [of ML] to humans”
- Kim & Doshi-Valez 2017
to left, formal definition (I)
that extends beyond humans
- Dhurandhar et al. 2017
10. Linear regression models
(with a certain number of parameters)
Decision trees (or similar)
(with a certain depth/number of parameters)
Text explanations
Visualizations (e.g., saliency maps)
more
Explanations come in many forms
12. from “The Mythos of Model Interpretability”, Lipton 2016
1. Simulatability, comprehend the entire model
at once; model complexity
2. Decomposability, comprehend the individual
components/parameters; intelligibility [2]
3. Algorithmic transparency, comprehend
the algorithm behavior; loss surface and randomness
Three levels of transparency
14. Interpretable methods survey
heavily borrowed from Kim & Doshi-Valez ICML 2017 tutorial
1. Fitting new models that are
intrinsically interpretable
2. Post-hoc analysis of existing model
3. Interpretable analysis of raw data
(or model architecture)
15. Inherently interpretable models
1. Fitting new models that are
intrinsically interpretable
Decision trees, rule lists, rule sets
Generalized linear models (and feature manipulation)
Case-based methods
Sparsity-based methods
Monotonicity-based methods
Conceptual and hierarchical models
16. 1. Fitting new models that are
intrinsically interpretable.
Decision trees, rule lists, rule sets
Generalized linear models (and feature manipulation)
table above from Gehrke et al. 2012
Inherently interpretable models
17. figure above from Gupta et al. 2016
Monotonicity constraints
Conceptual and hierarchical models
Inherently interpretable models
18. 2. Post-hoc analysis of existing model
Sensitivity analysis
Surrogate models
Gradient-based methods
Hidden layer investigations
Post-hoc for existing models
24. 3. Interpretable analysis of raw data
Visualization
Variable Importance
Partial dependence plots
Correlation analysis
Interpretable raw data analysis
27. Explanations can be persuasive
Lipton 2016, Herman 2017
When tailoring our model explanations to
human preferences and judges, our models
may learn to prioritize persuasive
explanations over introspective ones.
31. How AI detectives are cracking open the black box of
deep learning, Science Magazine, July 2017
Background reading
32. Ideas on interpreting machine learning,
O’Reilly Ideas, March 2017
Background reading
33. Thank you!
Let’s discuss this more.
Bernease Herman
bernease@uw.edu
@bernease on Twitter, Github, MSDSE Slack, everything
34. Splitting model form from simplicity
Herman 2017
Simultaneously
coerced into suitable
model form
(e.g., decision tree)
and reduced in
complexity
(e.g., model size).
Difficult to evaluate
across complexity
preferences.
35. Splitting model form from simplicity
Herman 2017
Keeps model form
and reduction of
complexity separate.
Improves evaluation
and adaptability.