When Evolutionary Computation Meets Trustworthy Artificial Intelligence
Xin Yao
School of Data Science, Lingnan University
Hong Kong SAR
Where is Lingnan University?
2
3
1. Trustworthy AI: What and Why
2. Fairer ML Through Multi-objective Evolutionary Learning
3. Multi-objective Feature Attribution Explanation
4. Concluding Remarks
Outline
4
1. Trustworthy AI: What and Why
2. Fairer ML Through Multi-objective Evolutionary Learning
3. Multi-objective Feature Attribution Explanation
4. Concluding Remarks
Outline
Trust, trustworthy systems and trustworthiness are not new words in science and engineering.
5
Six Attributes of Trustworthy Systems
In order to build trustworthy systems, we need to consider at least six attributes:
J. Li, B. Mao, Z. Liang, Z. Zhang, Q. Lin and X. Yao, "Trust and Trustworthiness: What They Are and How to Achieve Them," 2021 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops), Kassel, Germany, 2021, pp. 711-717, doi: 10.1109/PerComWorkshops51409.2021.9430929.
6
Seven System Requirements of Trustworthy Systems
J. Li, B. Mao, Z. Liang, Z. Zhang, Q. Lin and X. Yao, "Trust and Trustworthiness: What They Are and How to Achieve Them," 2021 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops), Kassel, Germany, 2021, pp. 711-717, doi: 10.1109/PerComWorkshops51409.2021.9430929.
7
From Trustworthy Systems to Trustworthy AI
Trustworthy AI has been a hot topic recently.
8
Examples of Trustworthy AI (I): Safety and Discrimination
9
Examples of Trustworthy AI (II): Robustness and Fairness
10
Examples of Trustworthy AI (III): Accuracy and Equality
11
What Does Trustworthy AI Should Include?
12
What Does Trustworthy AI Should Include?
13
What Does Trustworthy AI Should Include?
14
What Are Ethical Guidelines and Principles for AI?
In recent years, ethical issues related to AI have aroused widespread awareness and consideration among researchers, developers, users, enterprises, and governments.
15
Year | 2015 | 2016 | 2017 | 2018 | 2019 | 2020 | 2021 |
No. of Documents (146) | 2 | 7 | 25 | 53 | 31 | 24 | 4 |
To guide their strategies in developing, adopting, and embracing AI technologies, many organizations, including governments, companies, academic associations, and other national/international organizations, have established ethical frameworks or guidelines for the planning, development, production and usage of AI technology.
Ethical Guidelines and Principles for AI: Sources
Classification of AI ethical guidelines according to their sources.
16
Ethical Principles Identified in Existing Guidelines
17
Ethical principle | Number of�documents | Included codes/words |
Transparency | 107 | Transparency, explainability, explicability, understandability, interpretability, communication, disclosure, showing |
Fairness and Justice | 107 | Justice, fairness, consistency, inclusion, equality, equity, (non-) bias, (non-) discrimination, diversity, plurality, accessibility, reversibility, remedy, redress, challenge, access and distribution |
Responsibility | 100 | Responsibility, accountability, liability, acting with integrity |
Non-maleficence | 81 | Non-maleficence, security, safety, harm, protection, precaution, prevention, integrity (bodily or mental), non-subversion, reliability, robustness |
Privacy | 73 | Privacy, personal or private information, data protection, |
Beneficence | 41 | Benefits, beneficence, well-being, peace, social good, common good, non-violence |
Freedom and Autonomy | 34 | Freedom, autonomy, consent, choice, self-determination, liberty, empowerment, human rights |
Solidarity | 20 | Solidarity, social security, cohesion, inclusion, inclusiveness |
Sustainability | 19 | Sustainability, environment, nature, energy, resources |
Trust | 14 | Trust, trustworthiness, trustworthy |
Dignity | 13 | Dignity |
AI Ethics Is Hierarchical and Multi-dimensional
18
AI Ethics
Individual
Safety
Privacy & Data Protection
Freedom & Autonomy
Human Dignity
Group and Society
Fairness & Justice
Responsibility & Accountability
Transparency
Surveillance & Datafication
Controllability of AI
Democracy and Civil Rights
Job Replacement
Human Relationship
Environment and the Earth
Natural Resources
Energy
Environmental Protection
C. Huang Z. Zhang, B. Mao and X. Yao,“An Overview of Artificial Intelligence Ethics.”IEEE Transactions on Artificial Intelligence, vol. 4, no. 4, pp. 799-819, Aug. 2023, doi: 10.1109/TAI.2022.3194503.
Ethics and Trustworthiness in the Life Cycle of AI Systems
19
Stage of AI Lifecycle | Ethical Considerations Exist Along the Stage |
Business Analysis | Transparency, Fairness (Does the designed AI product includes any variables, features, processes that are unreasonable, morally objectionable, or unjustifiable?), Responsibility & Accountability, Democracy & Civil Rights, Sustainability |
Data Engineering | Privacy (How to assure the data security and keep the private and sensitive information included in data set?), Transparency (How to make data collection procedures transparent to consumers?), Fairness (Are data properly representative, relevant, accurate, and generalizable ?), Democracy & Civil Rights (How will you enable end users to control use of their data?) |
ML Modeling | Transparency (Does the decision or inference process of the model can be understood ?), Safety (Accuracy, Reliability, Security, and Robustness of the model), Fairness (Are the model outputs show disparate results on different groups of people ?) |
Model Deployment | Privacy (Make sure that private information cannot be re-identified through the deployed model), Safety (How to ensure the safety of the deployed model , such malicious modification and attack ? ) |
Operation &�Monitoring | Privacy (Privacy should be guaranteed during the operation & monitoring process), Fairness (Does the AI product has discriminatory or inequitable impacts on peoples they affect?), Democracy & Civil Rights (Do not infringe civil rights or the users) |
Procedural Trustworthiness + Outcome Trustworthiness
Now that we know trustworthy AI a little, how can we enhance it?
20
Enhancing AI Trustworthiness
21
Major Approaches
Technological Approaches
Non-technological Approaches
Specific Technological Approaches
The exiting work mainly focuses on some major issues and principles. Other issues and principles are rarely studied.
22
Principle | Representative Research Topics or Directions |
Transparency | Explainable AI (XAI) or Interpretable AI |
Fairness & Justice | Fair AI |
Non-maleficence | Safe AI Secure AI Robust AI |
Responsibility & Accountability | Responsible AI |
Privacy | Confidential Computing Differential Privacy Federated or Distributed Learning |
23
1. Trustworthy AI: What and Why
2. Fairer ML Through Multi-objective Evolutionary Learning
3. Multi-objective Feature Attribution Explanation
4. Concluding Remarks
Outline
Fair Machine Learning Is a Hot Research Topic
24
Search on Web of Science by using keys of “fairness and bias in artificial intelligence” or “algorithmic bias” or “algorithmic fairness” or “fairness-aware machine learning” or “fairness in machine learning”. (Accessed on Nov. 3, 2022)
Many Metrics Have Been Proposed to Measure Fairness
There are over 20 fairness metrics proposed so far [1].
25
[1] S. Verma and J. Rubin, “Fairness definitions explained,” in 2018 IEEE/ACM International Workshop on Software Fairness (FairWare), IEEE, 2018, pp. 1–7.
Fair Machine Learning Introduces Conflict
There are two inherent conflicts: (1) between model accuracy and fairness, and (2) among different fairness metrics.
26
(1) Conflict between accuracy and fairness
(2) Conflict among multiple fairness metrics
Image source: Figure 4 from [17] T. Speicher, H. Heidari, N. Grgic-Hlaca, K. P. Gummadi, A. Singla, A. Weller, and M. B. Zafarl, “A unified approach to quantifying algorithmic unfairness: Measuring individual &group unfairness via inequality indices,” in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2018, pp. 2239–2248
dotted line: unfairness
solid line: accuracy
dotted line: individual unfairness
solid line: group unfairness
Key Research Question
How can we make machine learning fairer according to different fairness metrics without sacrificing accuracy too much?
27
Multi-objective Fair Machine Learning
Key Idea: Treating the model accuracy and different fairness metrics as separate objectives in multi-objective learning [1].
Two studies were carried out：
28
[1] Q. Zhang, J. Liu, Z. Zhang, J. Wen, B. Mao and X. Yao, "Mitigating Unfairness via Evolutionary Multi-objective Ensemble Learning," in IEEE Transactions on Evolutionary Computation, vol. 27, no. 4, pp. 848-862, Aug. 2023, doi: 10.1109/TEVC.2022.3209544.
Study 1
Study 2
Multi-objective Fairer Machine Learning Framework
29
Q. Zhang, J. Liu, Z. Zhang, J. Wen, B. Mao and X. Yao, "Mitigating Unfairness via Evolutionary Multi-objective Ensemble Learning," in IEEE Transactions on Evolutionary Computation, vol. 27, no. 4, pp. 848-862, Aug. 2023, doi: 10.1109/TEVC.2022.3209544.
Initial population
Model training
Evaluate each model in population using metrics
Select better offspring models to replace some models in the population
Mating selection
Reproduction strategy
Stop?
……
Model 1
Model 2
Offspring
Model training
Evaluate each model in offspring using metrics
New Model 1
New Model 2
……
Yes
No
Output models in the final population
Does the Idea Work?
30
Latest Progress: Fairness-Aware Multiobjective Evolutionary Learning
Instead of using a fixed set of objectives during, we can let the algorithm to decide dynamically the set of objectives used in each generation of multi-objective evolutionary learning [1].
31
32
1. Trustworthy AI: What and Why
2. Fairer ML Through Multi-objective Evolutionary Learning
3. Multi-objective Feature Attribution Explanation
4. Concluding Remarks
Outline
What Is Feature Attribution Explanation in XAI？
Local feature attribution explanation (FAE) describes how much each input feature contributes to the output of a model for a given data point.
33
FAE explains a tabular data point [1]
FAE explains an image data point [2]
[1] Ribeiro M T , Singh S , Guestrin C . "Why Should I Trust You?": Explaining the Predictions of Any Classifier[C]// Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations. 2016.
[2] Lundberg S M, Lee S I. A unified approach to interpreting model predictions[C]//Proceedings of the 31st international conference on neural information processing systems. 2017: 4768-4777.
How to Evaluate Different Explanation Methods?
34
[1] Bhatt U , Weller A , Moura J . Evaluating and Aggregating Feature-based Model Explanations[C]// Twenty-Ninth International Joint Conference on Artificial Intelligence and Seventeenth Pacific Rim International Conference on Artificial Intelligence IJCAI-PRICAI-20. 2020.
Key Research Question
How can we find explanations that (a) balance tradeoff among different performance metrics, and (b) satisfy different explanation needs from different stakeholders?
35
Multi-Objective Feature Attribution Explanation (MOFAE)
Key Idea: Treating each evaluation metric as a separate objective in multi-objective feature attribution explanation.
36
Z. Wang, C. Huang, Y. Li and X. Yao, “Multi-objective Feature Attribution Explanation For Explainable Machine Learning,” ACM Transactions on Evolutionary Learning and Optimization, , Volume 4, Issue 1, Article No. 2, pp. 1–32. February 2024.. https://doi.org/10.1145/3617380.
Multi-Objective Feature Attribution Explanation (MOFAE): Framework
37
Initial population
Stop?
……
Offspring
……
Yes
No
Output final population
Does the Idea Work?
Yes, it worked well. We have answered positively the following three research questions in our study.
Q1. Do faithfulness, sensitivity, and complexity conflict with each other?
Q2. Can MOFAE simultaneously optimize these conflicting metrics and be competitive against existing state-of-the-art FAE methods?
Q3. Can our method find a set of explainable models (i.e., explanations) with different trade-offs among the objectives (i.e., metrics)?
38
Z. Wang, C. Huang, Y. Li and X. Yao, “Multi-objective Feature Attribution Explanation For Explainable Machine Learning,” ACM Transactions on Evolutionary Learning and Optimization, Volume 4, Issue 1, Article No. 2, pp. 1–32. February 2024. https://doi.org/10.1145/3617380.
Beyond FAE: A Roadmap for XAI
FAE is only a popular technique in XAI. There are more fundamental questions to be answered in XAI:
Explain to Whom, When, What and How? [1]
This question (actually questions) has not been fully answered by the existing literature!
39
[1] Z. Wang, C. Huang and X. Yao, “A Roadmap of Explainable Artificial Intelligence: Explain to Whom, When, What and How?,” ACM Transactions on Autonomous and Adaptive Systems, Volume 19, Issue 4, Article No. 20, Pages 1-40, November 2024, https://doi.org/10.1145/3702004.
To Whom: 9 Stakeholders
40
What to Explain
(1) How Explanations: How does the system work as a whole? For example, how does the autonomous driving system work?
(2) Why Explanations: Why does the system make a particular decision? For example, why did the car decide to slow down at this specific moment?
(3) Why-not Explanations: Why does the system not make a particular decision? For example, why did the vehicle not take an alternate route when encountering heavy traffic?
(4) What Explanations: What happens inside the system? For example, what has the AI system learned internally about recognizing obstacles?
(5) What-if Explanations: What would the system do if the input changes? For example, what if a pedestrian suddenly steps in front of the autonomous car, how would it react?
(6) What-else Explanations: What else are the similar instances? For example, what else are similar scenarios where autonomous vehicles also make decisions to turn?
(7) How-to Explanations: How to let the system make another particular decision? For example, how to change the environment to allow autonomous vehicles to choose a faster route?
(8) How-still Explanations: How much of a perturbation can there be while maintaining the same decision? For example, how much variation in road conditions can the vehicle withstand while still making the same decisions?
(9) Data Explanations: Ask for information about the data. For example, what biases exist in the training data of the autonomous vehicle system?
41
How to Explain
(1) Decision Trees (DTs) and Decision Rules (DRs).
(2) Feature Attribution Explanation (FAE).
(3) Partial dependence plot (PDP).
(4) Counterfactual (CF).
(5) Prototype Explanation (PE).
(6) Text Explanation (TE).
(7) Model Visualization (MV).
(8) Graph Explanation (GE).
(10) Exploratory Data Analysis (EDA).
42
When to Explain
AI System Lifecycle under Consideration
43
44
45
1. Trustworthy AI: What and Why
2. Fairer ML Through Multi-objective Evolutionary Learning
3. Multi-objective Feature Attribution Explanation
4. Concluding Remarks
Outline
Concluding Remarks
Email: xinyao@ln.edu.hk
46
Hong Kong PhD Fellowship Scheme (HKPFS)
The Fellowship provides
for a period up to three years.
Contact: xinyao@ln.edu.hk
RGC JRFS and LU PPRF