1. Challenges in Deploying
Explainable Machine Learning
Umang Bhatt
PhD Student, University of Cambridge
Research Fellow, Partnership on AI
Student Fellow, Leverhulme Center for the Future of Intelligence
@umangsbhatt

2. This Talk
1. How are existing approaches to explainability
used in practice?
2. Can existing explainability tools be used to ensure
model unfairness?
3. How can we create explainability tools for external
stakeholders?

3. This Talk
1. How are existing approaches to explainability
used in practice?
2. Can existing explainability tools be used to ensure
model unfairness?
3. How can we create explainability tools for external
stakeholders?

4. Explainable Machine Learning
in Deployment
Umang Bhatt, Alice Xiang, Shubham Sharma, Adrian Weller, Ankur Taly,
Yuhuan Jia, Joydeep Ghosh, Ruchir Puri, José Moura, and Peter Eckersley
Appeared at the ACM Conference on Fairness, Accountability, and Transparency 2020
https://arxiv.org/abs/1909.06342

5. Growth of Transparency Literature
Many algorithms proposed to “explain” machine learning
model output
We study how organizations use these algorithms, if at all

6. Our Approach
30 minute to 2 hour semi-structured interviews
50 individuals from 30 organizations interviewed

7. Shared Language
• Transparency: Providing stakeholders with relevant information about
how the model works: this includes documentation of the training
procedure, analysis of training data distribution, code releases, feature-
level explanations, etc.
• Explainability: Providing insights into a model’s behavior for specific
datapoint(s)

8. Our Questions
• What type of explanations have you used (e.g., feature-based, sample-
based, counterfactual, or natural language)?
• Who is the audience for the model explanation (e.g., research scientists,
product managers, domain experts, or users)?
• In what context have you deployed the explanations (e.g., informing the
development process, informing human decision makers about the model,
or informing the end user on how actions were taken based on the
model’s output)?

9. Types of Explanations
Feature Importance Sample Importance Counterfactuals

10. Stakeholders
Executives Engineers End Users Regulators

11. Findings
1.Explainability is used for debugging internally
2.Goals of explainability are not clearly defined
within organizations
3.Technical limitations make explainability hard
to deploy in real-time

12. Findings
1.Explainability is used for debugging internally
2.Goals of explainability are not clearly defined
within organizations
3.Technical limitations make explainability hard
to deploy in real-time

13. Use Cases

14. Findings
1.Explainability is used for debugging internally
2.Goals of explainability are not clearly defined
within organizations
3.Technical limitations make explainability hard
to deploy in real-time

15. Establishing Explainability Goals
1
Identify
stakeholders
2
Engage
stakeholders
W
hatpurpose
w
ith
the
explanation
serve?
3
Devise
w
orkflow
How
w
illthe
explanation
be
used
in
practice?
W
ho
w
illconsum
e
the
explanation?

16. Findings
1.Explainability is used for debugging internally
2.Goals of explainability are not clearly defined
within organizations
3.Technical limitations make explainability hard
to deploy in real-time

17. Limitations
• Spurious correlations exposed by feature level
explanations
• No causal underpinnings to the models themselves

18. Limitations (cont.)
• Sample importance is computationally infeasible to
deploy at scale
• Privacy concerns of model inversion exist

19. Findings
1.Explainability is used for debugging internally
2.Goals of explainability are not clearly defined
within organizations
3.Technical limitations make explainability hard
to deploy in real-time

20. This Talk
1. How are existing approaches to explainability
used in practice?
2. Can existing explainability tools be used to ensure
model unfairness?
3. How can we create explainability tools for external
stakeholders?

21. This Talk
1. How are existing approaches to explainability
used in practice? Only used by developers
2. Can existing explainability tools be used to ensure
model unfairness?
3. How can we create explainability tools for external
stakeholders?

22. This Talk
1. How are existing approaches to explainability
used in practice? Only used by developers
2. Can existing explainability tools be used to ensure
model unfairness?
3. How can we create explainability tools for external
stakeholders?

23. You Shouldn’t Trust Me: Learning
Models Which Conceal Unfairness
From Multiple Explanation Methods
Botty Dimanov, Umang Bhatt, Mateja Jamnik, and Adrian Weller
To appear at the European Conference on Artificial Intelligence 2020
http://ecai2020.eu/papers/72_paper.pdf

24. Takeaway
Feature importance reveals nothing reliable about model fairness.

25. Why do we care?
FeatureImportance
Feature
FeatureImportance
Feature
Model A Model B

26. Why do we care?
FeatureImportance
Feature
FeatureImportance
Feature
Model A Model B
Unfair Fair

27. Can we manipulate explanations?
Modified explanations via adversarial perturbations of inputs
• Ghorbani, Abid, and Zou. AAAI 2019
• Dombrowski et al. NeurIPS 2019
• Slack et al. AIES 2019
Control visual explanations via adversarial perturbations of parameters
• Heo, Joo, and Moon. NeurIPS 2019
Downgrade explanations via adversarial
perturbations of parameters to hide unfairness

28. fθ fθ+δ
f : X ↦ Y
Explanation g(f, x)j
Our Setup
Classifier Our Goal
∀i, fθ+δ(x(i)
) ≈ fθ(x(i)
)
Desirable Properties
Model Similarity
∀i, |g(fθ+δ, x(i)
)j | ≪ |g(fθ, x(i)
)j |Low Target Feature Attribution

29. Our Method
argminδ L′ = L(fθ+δ, x, y) +
α
n
∇X:,j
L(fθ+δ, x, y)
p
Adversarial Explanation Attack

30. Results (Importance Ranking)
fθ fθ+δ

31. Results (Importance Ranking)
fθ fθ+δ
Our adversarial explanation
attack:
1.Significantly decreases
relative importance.
2.Generalizes to test points.
3.Transfers across explanation
methods.

32. Findings
• Little change in accuracy, but
difference in outputs is detectable
• Low attribution achieved with
respect to multiple explanation
methods
• High unfairness across multiple
fairness metrics (compared to
holding feature constant)
fθ fθ+δ

33. This Talk
1. How are existing approaches to explainability
used in practice? Only used by developers
2. Can existing explainability tools be used to ensure
model unfairness?
3. How can we create explainability tools for external
stakeholders?

34. This Talk
1. How are existing approaches to explainability
used in practice? Only used by developers
2. Can existing explainability tools be used to ensure
model unfairness? Not in their current form
3. How can we create explainability tools for external
stakeholders?

35. This Talk
1. How are existing approaches to explainability
used in practice? Only used by developers
2. Can existing explainability tools be used to ensure
model unfairness? Not in their current form
3. How can we create explainability tools for external
stakeholders?

36. Machine Learning Explainability
for External Stakeholders
Umang Bhatt, McKane Andrus, Adrian Weller, and Alice Xiang
Convening hosted by CFI, PAI, and IBM
Appeared at the ICML 2020 Workshop on Extending Explainable AI: Beyond Deep Models and Classifiers
https://arxiv.org/abs/2007.05408

37. Overview
• 33 participants from 5 countries
• 15 ML experts, 3 designers, 6 legal experts, 9 policymakers
• Domain expertise: Finance, Healthcare, Media, and Social Services
• Goal: facilitate an inter-stakeholder conversation around explainable
machine learning

38. Two Takeaways
1. There is a need for community engagement in the development of
explainable machine learning.
2. There are nuances in the deployment of explainable machine learning.

39. Community Engagement
1. In which context will this explanation be used? Does the context change
the properties of the explanations we expose?
2. How should the explanation be evaluated? Both quantitatively and
qualitatively…
3. Can we prevent data misuse and preferential treatment by involving
affected groups in the development process?
4. Can we educate external stakeholders (and data scientists) regarding the
functionalities and limitations of explainable machine learning?

40. Deploying Explainability
1. How does uncertainty in the model and introduced by the (potentially
approximate) explanation technique affect the resulting explanations?
2. How can stakeholders interact with the resulting explanations? Can
explanations be a conduit for interacting with the model?
3. How, if at all, will stakeholder behavior change as a result of the
explanation shown?
4. Over time, how will the explanation technique adapt to changes in
stakeholder behavior?

41. Two Takeaways
1. There is a need for community engagement in the development of
explainable machine learning.
2. There are nuances in the deployment of explainable machine learning.

42. This Talk
1. How are existing approaches to explainability
used in practice? Only used by developers
2. Can existing explainability tools be used to ensure
model unfairness? Not in their current form
3. How can we create explainability tools for external
stakeholders?

43. This Talk
1. How are existing approaches to explainability
used in practice? Only used by developers
2. Can existing explainability tools be used to ensure
model unfairness? Not in their current form
3. How can we create explainability tools for external
stakeholders? Community engagement and
thoughtful deployment

44. This Talk
1. How are existing approaches to explainability
used in practice? Only used by developers
2. Can existing explainability tools be used to ensure
model unfairness? Not in their current form
3. How can we create explainability tools for external
stakeholders? Community engagement and
thoughtful deployment

45. Challenges in Deploying
Explainable Machine Learning
Umang Bhatt
usb20@cam.ac.uk
@umangsbhatt
Thanks for listening!