This book introduces a Responsible AI framework and guides you through processes to apply at each stage of the machine learning (ML) life cycle, from problem definition to deployment, to reduce and mitigate the risks and harms found in artificial intelligence (AI) technologies. AI offers the ability to solve many problems today if implemented correctly and responsibly. This book helps you avoid negative impacts – that in some cases have caused loss of life – and develop models that are fair, transparent, safe, secure, and robust.
The approach in this book raises your awareness of the missteps that can lead to negative outcomes in AI technologies and provides a Responsible AI framework to deliver responsible and ethical results in ML. It begins with an examination of the foundational elements of responsibility, principles, and data. Next comes guidance on implementation addressing issues such as fairness, transparency, safety, privacy, and robustness. The book helps you think responsibly while building AI and ML models and guides you through practical steps aimed at delivering responsible ML models, datasets, and products for your end users and customers.
What You Will Learn
• Build AI/ML models using Responsible AI frameworks and processes
• Document information on your datasets and improve data quality
• Measure fairness metrics in ML models
• Identify harms and risks per task and run safety evaluations on ML models
• Create transparent AI/ML models
• Develop Responsible AI principles and organizational guidelines
Who This Book Is For
AI and ML practitioners looking for guidance on building models that are fair, transparent, and ethical; those seeking awareness of the missteps that can lead to unintentional bias and harm from their AI algorithms; policy makers planning to craft laws, policies, and regulations that promote fairness and equity in automated algorithms
Author(s): Toju Duke
Publisher: Apress
Year: 2023
Language: English
Pages: 196
Table of Contents
About the Author
About the Technical Reviewer
Introduction
Part I: Foundation
Chapter 1: Responsibility
Avoiding the Blame Game
Being Accountable
Eliminating Toxicity
Thinking Fairly
Protecting Human Privacy
Ensuring Safety
Summary
Chapter 2: AI Principles
Fairness, Bias, and Human-Centered Values
Google
The Organisation for Economic Cooperation and Development (OECD)
The Australian Government
Transparency and Trust
Accountability
Social Benefits
Privacy, Safety, and Security
Summary
Chapter 3: Data
The History of Data
Data Ethics
Ownership
Data Control
Transparency
Accountability
Equality
Privacy
Intention
Outcomes
Data Curation
Best Practices
Annotation and Filtering
Rater Diversity
Synthetic Data
Data Cards and Datasheets
Model Cards
Tools
Alternative Datasets
Summary
Part II: Implementation
Chapter 4: Fairness
Defining Fairness
Equalized Odds
Equal Opportunity
Demographic Parity
Fairness Through Awareness
Fairness Through Unawareness
Treatment Equality
Test Fairness
Counterfactual Fairness
Fairness in Relational Domains
Conditional Statistical Parity
Types of Bias
Historical Bias
Representation Bias
Measurement Bias
Aggregation Bias
Evaluation Bias
Deployment Bias
Measuring Fairness
Fairness Tools
Summary
Chapter 5: Safety
AI Safety
Autonomous Learning with Benign Intent
Human Controlled with Benign Intent
Human Controlled with Malicious Intent
AI Harms
Discrimination, Hate Speech, and Exclusion
Information Hazards
Misinformation Harms
Malicious Uses
Human-Computer Interaction Harms
Environmental and Socioeconomic Harms
Mitigations and Technical Considerations
Benchmarking
Summary
Chapter 6: Human-in-the-Loop
Understanding Human-in-the-Loop
Human Annotation Case Study: Jigsaw Toxicity Classification
Rater Diversity Case Study: Jigsaw Toxicity Classification
Task Design
Measures
Results and Conclusion
Risks and Challenges
Summary
Chapter 7: Explainability
Explainable AI (XAI)
Implementing Explainable AI
Data Cards
Model Cards
Open-Source Toolkits
Accountability
Dimensions of AI Accountability
Governance Structures
Data
Performance Goals and Metrics
Monitoring Plans
Explainable AI Tools
Summary
Chapter 8: Privacy
Privacy Preserving AI
Federated Learning
Digging Deeper
Differential Privacy
Differential Privacy and Fairness Tradeoffs
Summary
Chapter 9: Robustness
Robust ML Models
Sampling
Bias Mitigation (Preprocessing)
Data Balancing
Data Augmentation
Cross-Validation
Ensembles
Bias Mitigation (In-Processing and Post-Processing)
Transfer Learning
Adversarial Training
Making Your ML Models Robust
Establish a Strong Baseline Model
Use Pretrained Models and Cloud APIs
Use AutoML
Make Model Improvements
Model Challenges
Data Quality
Model Decay
Feature Stability
Precision versus Recall
Input Perturbations
Summary
Part III: Ethical Considerations
Chapter 10: AI Ethics
Ethical Considerations for Large Language Models
Prevalent Discriminatory Language in LLMs
Working with Crowdworkers
Inequality and Job Quality
Impact on Creatives
Disparate Access to Language Model Benefits
Ethical Considerations for Generative Models
Deepfake Generation
Truthfulness, Accuracy, and Hallucinations
Copyright Infringement
Ethical Considerations for Computer Vision
Issues of Fraud
Inaccuracies
Consent Violations
Summary
Appendix A: References
Chapter 1
Chapter 2
Chapter 3
Chapter 4
Chapter 5
Chapter 6
Chapter 7
Chapter 8
Chapter 9
Chapter 10
Index
