EXPLAINABLE ARTIFICIAL INTELLIGENCE
School of Computing and Information Systems
Co-Director, Centre for AI & Digital Ethics The University of Melbourne @tmiller_unimelb
Copyright By PowCoder代写 加微信 powcoder
This material has been reproduced and communicated to you by or on behalf of the University of Melbourne pursuant to Part VB of the Copyright Act 1968 (the Act).
The material in this communication may be subject to copyright under the Act.
Any further copying or communication of this material by you may be the subject of copyright protection under the Act.
Do not remove this notice
LEARNING OUTCOMES
Motivate the need for explainable AI
Discuss the challenges of explainable AI
Describe the different types of explainable AI approaches
Describe the foundational approaches to explainable AI
RELATED READING
REQUIRED READING
“But why?” Understanding explainable artificial intelligence. .
XRDS 25, 3 (Spring 2019), 20–25. https://doi.org/10.1145/3313107 FURTHER READING FOR THOSE INTERESTED
Principles and Practice of Explainable Machine Learning. Vaishak Belle and . https://arxiv.org/pdf/2009.11698.pdf
This is an overview of explainability algorithms and research
Interpretable machine learning. . https://christophm.github.io/interpretable-ml-book/
A brilliant e-book on interpretable machine learning that is constantly improving
WHY ASK WHY?
WHY DO WE CARE ABOUT EXPLAINABILITY?
Source: K. Stubbs et al.: Autonomy and Common Ground in Human-Robot Interaction: A Field Study, IEEE Intelligent Systems, 22(2):42-50, 2007.
WHAT IS EXPLAINABLE AI?
Explainability = understanding
Explainable AI = understanding AI models and decisions
Explanation answers a why question
WHO CARES ABOUT EXPLAINABLE AI? WHY? WHEN?
How does a model What is driving Can I trust the work? decisions? model?
Data Scientist
Understand the model
Debug the model
Improve model performance
Business Owner
Understand the model
Evaluate model suitability
Accept model use
Risk Modeller
Challenge the model
Ensure model robustness
Approve model
Check impact of model on consumers
Verify model reliability
“What is the impact on me?”
“What actions can I take?”
Source: V. Belle and I. Papantonis: Principles and practice of explainable machine learning, arXiv, 2020. https://arxiv.org/abs/2009.11698
THE CHALLENGES OF
EXPLAINABLE AI
CHALLENGE: OPACITY
if respiratory_illness = Yes and smoker = Yes
and age >= 50 then lung_cancer
elif risk_lung_cancer = Yes and blood_pressure >= 0.3
then lung_cancer
elif risk_depression = Yes
and past_depression = Yes then depression
elif BMI >= 0.3
and insurance = None
and blood_pressure >= 0.2
then depression elif smoker = Yes
and BMI >= 0.2
and age >= 60 then diabetes
elif risk_diabetes = Yes and BMI >= 0.4
and prob_infections >= 0.2 then diabetes
Input layer
Output layer
Hidden layer 2
Hidden layer 1
CHALLENGE: CAUSALITY
“You are high risk of lung cancer because you have a high weekly alcohol intake”
CHALLENGE: CAUSALITY
Lung Cancer
Alcoholism
CORRELATED
CHALLENGE: CAUSALITY
CHALLENGE: THE HUMAN PROBLEM
PROPERTIES OF
EXPLAINABLE AI APPROACHES
GLOBAL VS. LOCAL
INTRINSIC VS. POST-HOC
No. wings = 2
true Spider
No. eyes≤4
No. legs ≥8
true Mozzie
false vs. Beetle
INTRINSIC AND POST-HOC
Source: Stiglic G, Kocbek S, Pernek I, Kokol P: Comprehensive Decision Tree Models in Bioinformatics. PLoS ONE 7(3): e33812, 2012.
MODEL-AGNOSTIC VS. MODEL-SPECIFIC MODEL SPECIFIC
Inner workings of model used for XAI
MODEL AGNOSTIC
Uses only inputs and outputs for XAI
Applicable to any model with that interface
FOUNDATIONAL METHODS IN EXPLAINABLE AI
ATTRIBUTION-BASED EXPLANATIONS
Source: T. Miller: “But why?” Understanding explainable artificial intelligence. XRDS 25, 3 (Spring 2019), 20–25. https://doi.org/10.1145/3313107
ATTRIBUTION-BASED EXPLANATIONS
Source: Ribeiro et al.: Why should I trust you?: Explaining the predictions of any classifier. In SIGKDD international conference on knowledge discovery and data mining. ACM, 2016.
EXAMPLE-BASED EXPLANATION: PROTOYPES
Source: Kim et al.: Examples are not enough, learn to criticize! Criticism for interpretability. In NeurIPS. 2016.
RULED-BASED EXPLANATION
EXTRACT RULES POST-HOC
LEARN INTERPRETABLE RULES DIRECTLY
if respiratory_illness = Yes and smoker = Yes
and age >= 50 then lung_cancer
elif risk_lung_cancer = Yes and blood_pressure >= 0.3
then lung_cancer
elif risk_depression = Yes
and past_depression = Yes then depression
elif BMI >= 0.3
and insurance = None
and blood_pressure >= 0.2
then depression elif smoker = Yes
and BMI >= 0.2
and age >= 60 then diabetes
elif risk_diabetes = Yes and BMI >= 0.4
and prob_infections >= 0.2 then diabetes
EXPLANATIONS ARE
CONTRASTIVE
“The key insight is to recognise that one does not explain events per se, but that one explains why the puzzling event occurred in the target cases but not in some counterfactual contrast case”
DENIS HILTON: CONVERSATIONAL PROCESSES AND CAUSAL EXPLANATION, PSYCHOLOGICAL BULLETIN. 107(11):65-81, (1990)
CONTRASTIVE EXPLANATION THE DIFFERENCE CONDITION
Type Legs Stinger Eyes Eyes Wings
Spider 8 ✘ 8 ✘ 0
Beetle 6 ✘ 2 ✔ 2 FLY?
Bee 6 ✔ 5 ✔ 4 Fly 6 ✘ 5 ✔ 2
WHY IS IT A
CONTRASTIVE EXPLANATION THE DIFFERENCE CONDITION
Type Legs Stinger Eyes Eyes Wings
Spider 8 ✘ 8 ✘ 0
Beetle 6 ✘ 2 ✔ 2 FLY?
Bee 6 ✔ 5 ✔ 4 Fly 6 ✘ 5 ✔ 2
WHY IS IT A
CONTRASTIVE EXPLANATION THE DIFFERENCE CONDITION
Type Legs Stinger Eyes Eyes Wings
Spider 8 ✘ 8 ✘ 0
Beetle 6 ✘ 2 ✔ 2 FLY?
Bee 6 ✔ 5 ✔ 4 Fly 6 ✘ 5 ✔ 2
WHY IS IT A
CONTRASTIVE EXPLANATION THE DIFFERENCE CONDITION
Type Legs Stinger Eyes Eyes Wings
Spider 8 ✘ 8 ✘ 0
WHY IS IT A FLY Beetle 6 ✘ 2 ✔ 2 RATHER THAN A
Bee 6 ✔ 5 ✔ 4 Fly 6 ✘ 5 ✔ 2
CONTRASTIVE EXPLANATION THE DIFFERENCE CONDITION
Type Legs Stinger Eyes Eyes Wings
Spider 8 ✘ 8 ✘ 0
WHY IS IT A FLY Beetle 6 ✘ 2 ✔ 2 RATHER THAN A
Bee 6 ✔ 5 ✔ 4 Fly 6 ✘ 5 ✔ 2
CONTRASTIVE EXPLANATION THE DIFFERENCE CONDITION
Type Legs Stinger Eyes Eyes Wings
Spider 8 ✘ 8 ✘ 0
WHY IS IT A FLY Beetle 6 ✘ 2 ✔ 2 RATHER THAN A
Bee 6 ✔ 5 ✔ 4 Fly 6 ✘ 5 ✔ 2
EXPLAINABLE AI: SUMMARY
EXPLAINABILITY
Different people with different explainability needs
IMPROVE DECISION-MAKING ETHICS AND ACCOUNTABILITY
Both human and technical challenges
HUMAN INTERPRETATION
EXPLAINABILITY APPROACHES
Properties
LOCAL vs. GLOBAL
INTERPRETABLE vs. POST-HOC MODEL-AGNOSTIC vs. MODEL-SPECIFIC
Key approaches
ATTRIBUTION
CONTRASTIVE EXPLANATION
School of Computing and Information Systems Co-Director, Centre for AI & Digital Ethics
The University of Melbourne
@tmiller_unimelb
程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com