まとめページへのリンク

Explainable AI/ML (XAI) for Accountability, Fairness, and Transparency
Page: https://xai.kdd2019.a.intuit.com/

Keynote: Responsible Use of Explainable Artificial Intelligence

Patrick Hall
https://github.com/jphall663/kdd_2019/

Intro
Understanding the trust
The dark side
Surrogates

Intro
What is an explanaition in ML?
What do I mean by explainable ML?
Mostly post-hoc techniques users to enhance understanding od trained model mechanism and prediction
Direct measures od global and local feature

Responsible use of explainable ML can enable

Human learning from machine learning
Human appeal of automated decision
Regulatory compliance
White-hat hacking and security audits of ML model
Even logistic regression is often “explained ",or post-processed for credit scoring, eg)Max points lost method and adverse action notices.

Why propose guidelines

Misuse and abuseof explainable ML can enable
ExplainableML is already in-use
Regulatory guidance is not agreed upon yet

Proposed guidelines for responsible use

Use explainable ML to enhance understaning
Learn how explainable ML is used for nefarious purpose.
Augment surrogate models with direct explanations
Use highly transparent machanism for high stakes application

Use explainable ML to enhance understaning

Use explainable ML to enhance understaning
corollary2.1: White-hat attackes can crack potentially harmful black-box
corollary2.2: Explanation is not a front line fairness tool
Surrogates models
BUT many explainable ML techniques have nothing to do with surrogate models

Augment surrogate models with direct explanations
corollary3.1: augment LIME with direct explanations

Use highly transparent machanism for high stakes application
corollary4.1: explanations and interpretable models are not mutually exclusive.
Interrude: an ode to the shaply values
corollary4.3: explanations and fairness techniques are not mutually exclusive

Contributed Talk: Parametrised Data Sampling for Fairness Optimisation

Carlos Vladimiro Gonzalez Zelaya, Paolo Missier, and Dennis Prangle

Slides: https://github.com/vladoxNCL/fairCorrect/blob/master/Presentation.pdf

outline

Method for correcting classifier fairness
Model and definition agonistic
Tune correction level to optimize fairness

Train set correlation
Sampling strategies

Original data
Under
SMOTE[chawia et al 2002]
Preferential
Equality form
Ratio form
Some fairness ratios
Demographic parity
Equality of opportunity
Counterfactual (proxy)
Effect on test set (COMPAS, Undersampling)
Effect is correlated with correction
But it occurs to a different effect
Intersection is not at d=0
Optional correction by fairness and sampling

Accuracy vs fairness trade-off
Conjecture: A LR model is FTU <> CFR = 1

Conclusion

Fairness-agonistic optimization with a relatively small loss in accuracy
ideal correction level is definition dependent
Different sampling strategy produced similar result

Contributed Talk: What Makes a Good Explanation? The Lemonade Stand Game As a Platform for Evaluating XAI
Talia Tron, Daniel Ben David, Tzvika Barenholz, and Yehezkel S. Resheff

Title: What can selling lemonade teach us about XAI
What to explain:
Nothing > the data > the model > the prediction
Model transparency
Global explanation
local explanation

What makes a good explanation
Explanation quality
Users understanding

user behavior adoption trust, satisfaction

The challenge

Behabioal measures
Dynamic
Real financial
Counrolled environment
Uncertain conditions

Research questions

Evaluate the influence of different explanations on financial alporism advice adoption

Experiment: the situation of selling lemonede.

Keynote: Interpretability - now what?

Been Kim
Sea of interpretability

Where are we going ?
What do we have now ?
What can we do better ?
What should we be carefull ?

We need to ensure to
1, Our values are aligned
2, Our knowledge is reflected for everyone.

Not out goal

About making all models interpretable
About understanding every single bit about the model
Against developing highly complex models
Only about gaining user trust of fairness

What do we have now ?
Investing post-training interpretability
Saliency Maps
Some confusing behavior of saliency Maps
Sanity checlk1
When prediction changes do explanations change ? NO
Network rained with true and random labels do explanations deliver different explanation

What can we learn from this?

Confirmation bias
Others who independently reached the same conclusions
Some of these methods have been shown to be useful for humans why ? More studyes are needed .
Resent work by Gupta Arora 19’ suggests a simple fix

Goal of TCAV: Testing with concept activation vectors
Defining concept activation vector (CAV)
TCAV core idea: derivative with CAV to get prediction sensitivity
Guiding against spurious CAV
Quantitive validation : guarding against spurious CAV
Recap TCAV: testing with concept activation vectors.

Limitation of TCAV
Concept ohas to “exoressible” Using examples
User needs to know which concepts they want to test
Explanation is Not-casual

What should we be carefull ?
Proper evaluation
Remember that humans are biasied and irrational
Design the right interaction (HCI)
Try to criticize (think about what wasn’t talked about in this talk but should have )
keep checking