Responsible AI at Scale (Sarah Bird, Microsoft)

It’s critical that AI systems reflect our values and principles. For organizations developing AI systems at scale, putting principles into practice poses unique challenges. In Azure AI, doing this effectively has required us to innovate across our entire product lifecycle: in our development processes; in the technology that we build; and in the way in which we make technology available. In this talk, I will share our experiences implementing responsible AI at scale across our speech, vision, language, and decision systems.

Transcript

1. Responsible AI at Scale Sarah Bird Azure AI
2. AI is Disruptive Fears include workforce displacement, loss of privacy, potentially unfair decision-making and lack of control over automated systems and robots. Accenture Responsible AI PDF
3. What is Responsible AI? While these issues are significant, they are also addressable with the right planning, innovation, and governance. Responsible AI is a framework for bringing many of these critical practices together. It focuses on ensuring the ethical, transparent and accountable use of AI technologies in a manner consistent with user expectations, organizational values and societal laws and norms.
4. MICROSOFT’S AI PRINCIPLES Reliability & Safety Transparency Inclusiveness Privacy & Security Fairness Accountability
5. Responsible AI Process Assessment Responsible development Responsible use Evolving
6. Responsible AI Process Assessment
7. Alex Kipman’s voice Custom Neural Voice trained with 200+ sentences Synthesized Recorded Synthesized (ja-JP) Custom Neural Voice
8. Protect listeners from misleading content and experiences Protect speakers from having their voices impersonated or improperly used
9. Trade-offs & tensions New dimensions we need to consider in product development
10. Responsible AI Process Assessment Responsible development
11. Identify priority factors and groups Assess system for disparities between groups Mitigate disparities between groups Reassess as part of regular testing Assessing and Mitigating Fairness
12. Factors Factors Speech-to-Text Facial Recognition Spatial Analysis Ancestry Y Y Y Gender Identity Y Y Y Age Y Y Y Accessibility: Mobility Aids N/A N/A Y Non-Native Accent Y N/A N/A Regional Dialects Y N/A N/A Ancestry x Gender Y Y Y Ancestry x Age Y Y Y Ancestry x Gender x Age Y Y Y Generally Applicable Service Specific Intersections
13. Priority Factors for Data Collection Factor with target distribution • Ancestry • Age • Gender identity • Geographic region • Living environment • Context of speech Screener or metadata • Education level • Household income • Location(s) language was learned • Native language(s)
14. RAI during data collection Data collection plan Data Source Collection protocol Annotation Collect Design Procure Label
15. Assessing unfairness in your model 1 Fairness assessment Use common fairness metrics and an interactive dashboard to assess which groups of people may be negatively impacted Model formats Python models using scikit predict convention, Scikit, Tensorflow, Pytorch, Keras Metrics 15+ common group fairness metrics Model types Classification, Regression 2 Unfairness Mitigation Use state-of-the-art algorithms to mitigate unfairness in your classification and regression models https://github.com/fairlearn/fairlearn
16. Understand and debug your model Interpret Glassbox and blackbox interpretability methods for tabular data Interpret-community Additional interpretability techniques for tabular data Interpret-text Interpretability methods for text data DiCE Diverse counterfactual explanations https://github.com/interpretml Blackbox models Model formats: Python models using scickit predict convention, Scikit, Tensorflow, Pytorch, Keras Explainers: SHAP, LIME, Global Surrogate, Feature Permutation Glassbox models Model types: Linear Models, Decision Trees, Decision Rules, Explainable Boosting Machines AzurML-interpret AzureML SDK wrapper for Interpret and Interpret-community
17. Identify and Diagnosis Errors Income Prediction UCI github.com/microsoft/responsible-ai-widgets Error Analysis
18. Responsible AI Process Assessment Responsible development Responsible use
19. Computer Vision for Spatial Analysis Enables building applications to analyze live video and understand people’s movement in physical space.
20. System limitations and best practices to improve system accuracy •Spatial analysis should not be relied on for scenarios where real-time alerts are needed to trigger intervention to prevent injury, like turning off a piece of heavy machinery when a person is present. Space analytics is better used to reduce the number of unsafe acts by measuring the aggregate number of people violating rules like entering restricted/forbidden areas. •Spatial analysis has not heavily tested with data containing minors under the age of 18 or adults over age 65. We would recommend that customers thoroughly evaluate error rates for their scenario in environments where these ages predominate. •Spatial analysis face mask detection attribute should not be relied on if a person is wearing a transparent shield or glittery face masks; they make it challenging for the system to function accurately. •Spatial analysis will work best when configured with a ~15 frames per second input video stream with at least 1080p resolution. A slower frame rate or lower resolution risks losing track of people when they move quickly or are too small in the camera view
21. Responsible AI documentation Spatial Analysis Transparency Note Describes how spatial analysis works, the intended uses and use cases to avoid, as well as key characteristics which impact performance and known limitations. Data, Privacy, and Security Describes what data spatial analysis processes, and how spatial analysis processes that data. It also provides links to information on the security measures provided by the container and supported services like Azure IoT Edge. Integration and Responsible Use Provides recommendations based on customer experience on how to uphold responsible AI principles in deployment. Specifically covers privacy preserving techniques, disclosure guidelines, and approaches to support effective human decision making. Is accompanied by design prototypes for disclosure and founding research. Responsible use deployment for Computer Vision spatial analysis - Azure Cognitive Services | Microsoft Docs
22. Custom Neural Voice Protect listeners from misleading content and experiences Protect speakers from having their voices impersonated or improperly used
23. Limited Access
24. Collecting Training Data – Consent from Voice Talents
25. Build a Voice – Select Consent Statement for Verification Selecting the wrong talent
26. Build a Voice – Verification Failed Failed due to Voice Talent Mismatch
27. Build a Voice – Select Consent Statement for Verification Selecting the correct talent
28. Build a Voice – Verification and Training Succeeded Succeeded with Talent Match
29. Integrate – Deployment and Use Requirements
30. Responsible AI Process Assessment Responsible development Responsible use Evolving
31. Learn More Responsible AI Resource Center https://www.microsoft.com/en-us/ai/responsible-ai-resources Azure AI https://azure.microsoft.com/en-us/overview/ai-platform/