“Identifying and Mitigating Bias in AI,” a Presentation from Intel
0 likes203 views
The document discusses the importance of responsible AI (RAI) and highlights the risks of bias in AI systems, including historical, representation, measurement, and deployment biases. It presents various tools and metrics for identifying and mitigating bias, such as IBM's AI Fairness 360 and Google's What-If Tool, emphasizing the need for transparency, human oversight, and ethical considerations in AI development. Key takeaways include adopting RAI principles, conducting regular audits, and keeping abreast of evolving regulations to foster trust and equity in AI applications.
“Identifying and Mitigating Bias in AI,” a Presentation from Intel
- 1. Identifying and Mitigating Bias in AI Nikita Tiwari AI Enabling Engineer, Client Computing Group Intel Corporation
- 2. • Why Should we Care About Responsible AI (RAI)? • Identifying Bias in AI • Tools to Mitigate Bias in AI • Fairness Metrics • What-If Tool • AI Fairness 360 • Key Takeaways • Industry-Wide Ethical AI Revolution & Resources Agenda 2 © 2024 Intel Corporation
- 3. 3 © 2024 Intel Corporation Discussion Ethics is a conversation! AI stories you have come across (positive & negative) Why should we care?
- 4. Why Responsible AI? 4 © 2024 Intel Corporation Harm to human life Loss of trust Fines in compliance & regulations Introduction of systemic bias Misinformation Breach of privacy AI Incidents in the News Barriers in implementing trustworthy and explainable AI Daily reports of AI harm Generative AI accessible to all end users Massive global AI market size Cost of AI Incidents
- 5. Defining Responsible AI Principles Respect Human Rights Enable Human Oversight Transparency Personal Privacy Security, Safety, Sustainability Equity and Inclusion Focus on data used for training and the algorithm development process to help prevent bias and discrimination. Understanding where the data came from and how the model works. Human oversight of AI solutions to ensure they positively benefit society. Ethical review and enforcement of end-to-end AI safety. Low- resource implementation of AI algorithms Maintaining personal privacy and consent. Focusing on protecting the collected data. Human rights are a cornerstone for AI development. AI solutions should not support or tolerate usages that violate human rights. © 2024 Intel Corporation 5
- 6. 6 © 2024 Intel Corporation Bias Identification
- 7. Historical Bias: • Caused by preconceived notion even on perfectly measured/sampled data • E.g., Gendered occupation • Case study: Amazon hiring algorithm/ Google Gemini AI image generation Bias in Data Generation 7 © 2024 Intel Corporation Representation Bias: • Target population does not: • reflect the user population • include underrepresented groups • Sampling method is limited or uneven • Two-fold representation – uniform vs proportional • E.g., ImageNet images Measurement Bias: • The proxy oversimplifies a complex construct, or measurement methods and accuracy differ among groups • E.g., COMPAS (Correctional Offender Management Profiling for Alternative Sanctions)
- 8. Aggregation Bias • Arises when data from different sources or groups are combined, leading to distortions in the model's performance or predictions • E.g., Housing price prediction model trained on data aggregated from multiple cities without accounting for differences in housing markets Bias in Model Building and Implementation (1/2) 8 Learning Bias • Arises when modeling choices amplify performance disparities across different examples in the data • E.g., Prioritizing one objective damages the other, like optimizing for privacy or compactness may affect accuracy © 2024 Intel Corporation
- 9. Evaluation Bias • The benchmark data or the evaluation process used for a particular task does not represent the user population or favors certain groups over the others • E.g., A language translation model evaluated primarily on its accuracy for European languages, hinders its usefulness in diverse contexts as a global language translator Bias in Model Building and Implementation (2/2) 9 Deployment Bias • Arises due to “off-the-label” usage of model other than intended use • E.g., Models intended to predict a person’s likelihood of committing a future crime also used to determine the length of the sentence © 2024 Intel Corporation
- 10. 10 © 2024 Intel Corporation Bias Mitigation
- 11. • Fairness metrics are a set of measures that enable you to detect the presence of bias in your data or model • At least 21 fairness metrics. Many are conflicting. Which is the fairest? No right answer. Some common metrics - • Group unaware – Removes all group and proxy-group membership information from the dataset to avoid favoring any groups or sub-class. Similar logic as unsupervised learning. Difficult to achieve. • Group threshold – Alternate thought-process to group unaware. Adjust the confidence thresholds for different groups independently such that the confidence threshold for correct predictions for a minority group will be slightly lower. • Demographic parity – Similar percentages of datapoints from each group are predicted as positive classifications. E.g., A class with x% of subclass-1 should have x% of subclass-1 positive predictions. • Equal opportunity – Among those datapoints with the positive ground truth label (true positive rate), there is a similar percentage of positive predictions in each group. • Equal accuracy – The model’s predictions are equally accurate for all groups. True positive and false positive should be same across all groups. Common Fairness Metrics 11 © 2024 Intel Corporation
- 12. What-If Tool (Google) • Simulation with data manipulation & specific criteria to detect bias • 5 fairness metrics • Bias mitigation not straightforward Tools to Detect and Mitigate Bias 12 AI Fairness 360 (IBM) • Extensible toolkit for bias detection & mitigation • 70+ fairness metrics • 10 bias mitigation algorithms • Fairness metric explanations © 2024 Intel Corporation • Key challenge in developing and deploying a ML system is understanding their performance across a wide range of inputs • Several open-source tools utilize fairness metrics and bias mitigation algorithms to analyze ML systems with limited coding and test bias in hypothetical scenarios
- 13. 13 © 2024 Intel Corporation Key Takeaways and Resources
- 14. • Establish RAI principles that guide the decision-making for your AI development • Drive RAI requirements into product definition • Adopt human-centric approach at every stage of your product development • Integrate RAI tools in your software development lifecycle • Preventing bias is complex. Define fairness metrics, document trade-offs and share with your users transparently. Re-check for bias often. • Conduct regular assessments, audits, and update AI response plans • Keep up-to-date with the evolving legislature, regional laws and standards • Certifications can help adherence to standards, legislatures and build user trust Key Takeaways 14 © 2024 Intel Corporation
- 15. • Responsible Artificial Intelligence Institute (RAII) • First independent, accredited certification program for RAI in US, Canada, Europe and UK • Vectors: Systems operations, explainability and interpretability, accountability, consumer protection, bias and fairness, and robustness with collaborations across world economic forum, OECD, IEEE, ANSI, etc., • Executive orders on responsible AI around the globe • European Union AI act • Executive Order on the Safe, Secure and Trustworthy Development and use of AI • The White House Blueprint for an AI Bill of Rights Industry-Wide Ethical AI Revolution 15 • IEEE Standards Association • https://standards.ieee.org/participate/ • 2100+ Standards, 175+ Countries, 34000+ Global Participants • IEEE CertifAIEd™ (part of Standards Association) • A certification program for assessing ethics of Autonomous Intelligent Systems (AIS) • Vectors: Ethical Privacy, Algorithmic Bias, Transparency, and Accountability, Agentic AI • AI Safety Certification/Assurance Development • Case Study for AI ethics applied to the city of Viennna © 2024 Intel Corporation
- 16. Resources 16 © 2024 Intel Corporation Responsible AI Landscape https://hai.stanford.edu/news/2022-ai-index-industrialization- ai-and-mounting-ethical-concerns Grandview Research https://www.grandviewresearch.com/industry- analysis/artificial-intelligence-ai-market Global AI adoption Index https://filecache.mediaroom.com/mr5mr_ibmnewsroom/191 468/IBM%27s%20Global%20AI%20Adoption%20Index%20202 1_Executive-Summary.pdf Measuring Bias – David Weinberger https://pair-code.github.io/what-if-tool/ai-fairness.html Intel Responsible AI Program https://www.intel.com/content/www/us/en/artificial- intelligence/responsible-ai.html Responsible AI Institute Certification https://www.responsible.ai/how-we-help IEEE CertifAIEd™ https://engagestandards.ieee.org/ieeecertifaied.html AI Incident database https://incidentdatabase.ai/ European Union AI Act https://ec.europa.eu/commission/presscorner/detail/en/IP_23_ 6473 White House Executive Order https://www.whitehouse.gov/briefing-room/presidential- actions/2023/10/30/executive-order-on-the-safe-secure-and- trustworthy-development-and-use-of-artificial-intelligence/ Blueprint for AI Bill of Rights https://www.whitehouse.gov/ostp/ai-bill-of-rights/
- 17. 17 © 2024 Intel Corporation Questions
- 18. 18 © 2024 Intel Corporation Backup Material
- 19. Designing With a Human-Centric Approach 19 © 2024 Intel Corporation Definition Does AI add value? Who are the indented users of the system? Identify intended potential harm and plan for remediations Translate user needs into data needs E.g., Prototyping a chatbot Development Source high-quality unbiased data responsibly Get inputs from domain experts Enable human oversight Built-in safety measures E.g., Improving autonomous vehicles Deployment Provide ways for users to challenge the outcome Provide manual controls when AI fails Offer high-touch customer support Marketing Focus on the benefit, not the technology Transparently share the limitations of the system with the users Be transparent about privacy and data settings Anchor on familiarity
- 20. Google Gemini AI Image Generation Mistake What happened? Why did the incident happen? Remediation and next steps https://blog.google/products/gemini/gemini-image-generation-issue/ © 2024 Intel Corporation Prabhakar Raghavan Senior Vice President, Google Feb 23, 2024 20
- 21. • An AI algorithm was used to recognize when person moves away from laptop and to turn off the screen • The algorithm was tested to be inclusive and performant on individuals with different skin tones, to ensure the algorithm is fair and the output is not affected. Intel Ethical AI Impact Assessment Case Studies 21 • Pedestrian detection for self-driving cars should incorporate diverse data, including data from disabled pedestrians, such as folks in wheelchairs © 2024 Intel Corporation Diverse Skin Tone Recognition with Intel Evo Laptop Pedestrian Detection including disabled individuals
- 22. Open-source library Notes AIF360 Provides a comprehensive set of metrics for datasets and models to test for biases and algorithms to mitigate bias in datasets and models. Fairness Measures Provides several fairness metrics, including difference of means, disparate impact, and odds ratio. It also provides datasets, but some are not in the public domain and require explicit permission from the owners to access or use the data. FairML Provides an auditing tool for predictive models by quantifying the relative effects of various inputs on a model’s predictions, which can be used to assess the model’s fairness. FairTest Checks for associations between predicted labels and protected attributes. The methodology also provides a way to identify regions of the input space where an algorithm might incur unusually high errors. This toolkit also includes a rich catalog of datasets Aequitas This is an auditing toolkit for data scientists as well as policymakers; it has a Python library and website where data can be uploaded for bias analysis. It offers several fairness metrics, including demographic, statistical parity, and disparate impact, along with a “fairness tree” to help users identify the correct metric to use for their particular situation. Aequitas’s license does not allow commercial use. Themis An open-source bias toolbox that automatically generates test suites to measure discrimination in decisions made by a predictive system. Themis-ML Provides fairness metrics, such as mean difference, some bias mitigation algorithms, additive counterfactually fair estimator, and reject option classification. Fairness Comparison Includes several bias detection metrics as well as bias mitigation methods, including disparate impact remover, prejudice remover, and two-Naive Bayes. Written primarily as a test bed to allow different bias metrics and algorithms to be compared in a consistent way, it also allows additional algorithms and datasets. Open-source Fairness Metrics Libraries 22 © 2024 Intel Corporation