Trustworthy AI and Open Science
0 likes281 views
The document discusses the intersection of trustworthy AI and open science, emphasizing the importance of visibility and accountability in research funded by the NSF. It outlines various types of data repositories and their roles in open science, highlighting the trade-offs between generalist and specialist repositories. Additionally, it addresses the need for trustworthy AI, focusing on developer ethics, explainability, and societal influences to ensure accountability in AI systems.
Trustworthy AI and Open Science
- 1. TRUSTWORTHY AI AND OPEN SCIENCE Beth Plale Michael A and Laurie Burns McRobbie Professor of Computer Engineering Beilstein Open Science symposium October 06, 2021 Luddy School of Informatics, Computing, and Engineering Data To Insight Center
- 2. Observations influenced by my role (2017-2020) in the National Science Foundation working on agency policies and practice in open science. Views expressed are entirely my own. Funding agency perspective on open science: how do we bring visibility to the products of research (that we fund)
- 3. NSF funds the collection and capture of research data through projects ranging from a few hundred thousand dollars to tens of millions of dollars. The data are maintained in a landscape of solutions to meet the needs of researchers.
- 4. Specialist repositories - Organizational resources Generalist repositories - Organizational resources Data Portals - Low velocity data - Employs cloud resources - Employs data-compute proximity for analysis Observation networks - High velocity data - Employs cloud resources RESEARCH DATA LANDSCAPE SAGE NEON ARM HPWREN UWI LTER, OOI NEON HydroShare LTER MGDS, IRIS ICPSR QDR TAIR MDF IEDA PDB CCDC DataVerse Figshare Dryad Zenodo IRs Exemplar systems
- 5. RESEARCH DATA LANDSCAPE Data timeliness need Researcher depth of expertise Expectation for level of curation Expectation of data longevity Specialist repositories - Organizational resources Generalist repositories - Organizational resources Data Portals - Low velocity data - Employs cloud resources - Employs data-compute proximity for analysis Observation networks - High velocity data - Employs cloud resources SAGE NEON ARM HPWREN UWI LTER, OOI NEON HydroShare LTER MGDS, IRIS ICPSR QDR TAIR MDF IEDA PDB CCDC DataVerse Figshare Dryad Zenodo IRs
- 6. RESEARCH DATA LANDSCAPE Publisher’s view of landscape (general public view as well) Optimization for timeliness of research could suggest lower value over time Specialist repositories - Organizational resources Generalist repositories - Organizational resources Data Portals - Low velocity data - Employs cloud resources - Employs data-compute proximity for analysis Observation networks - High velocity data - Employs cloud resources SAGE NEON ARM HPWREN UWI LTER, OOI NEON HydroShare LTER MGDS, IRIS ICPSR QDR TAIR MDF IEDA PDB CCDC DataVerse Figshare Dryad Zenodo IRs
- 7. Generalist–Aided Deposit: engages generalist curators Metadata: generalist schema Reuse potential: moderate-low as metadata is curated but general Scope: discipline agnostic scope Discovery: broad name recognition Specialist-DBMS Deposit: difficult so DB often read-only Metadata: data dictionary + DB schema Reuse potential: high potential as self contained Scope: subdiscipline scope Discovery: known within subdiscipline Specialist–Aided Deposit: engages specialist curators Metadata: specialized schema Reuse potential: high due to specialists Scope: discipline scope Discovery: known within discipline Specialist-Unaided Deposit: unaided deposit Metadata: specialized schema Reuse potential: moderate-high from discipline focus of metadata schema Scope: discipline scope Discovery: known within discipline Generalist-Unaided Deposit: unaided deposit Metadata: generalist schema Reuse potential: low as metadata is minimal Scope: discipline agnostic scope Discovery: broad name recognition i.e., institutional repositories
- 8. Generalist–Aided Deposit: engages generalist curators Metadata: generalist schema Reuse potential: moderate-low as metadata is curated but general Scope: discipline agnostic scope Discovery: broad name recognition Specialist-DBMS Deposit: difficult so DB often read-only Metadata: data dictionary + DB schema Reuse potential: high potential as self contained Scope: subdiscipline scope Discovery: known within subdiscipline Specialist–Aided Deposit: engages specialist curators Metadata: specialized schema Reuse potential: high due to specialists Scope: discipline scope Discovery: known within discipline Specialist-Unaided Deposit: unaided deposit Metadata: specialized schema Reuse potential: moderate-high from discipline focus of metadata schema Scope: discipline scope Discovery: known within discipline Generalist-Unaided Deposit: unaided deposit Metadata: generalist schema Reuse potential: low as metadata is minimal Scope: discipline agnostic scope Discovery: broad name recognition i.e., institutional repositories
- 9. Generalist–Aided Deposit: engages generalist curators Metadata: generalist schema Reuse potential: moderate-low as metadata is curated but general Scope: discipline agnostic scope Discovery: broad name recognition Specialist-DBMS Deposit: difficult so DB often read-only Metadata: data dictionary + DB schema Reuse potential: high potential as self contained Scope: subdiscipline scope Discovery: known within subdiscipline Specialist–Aided Deposit: engages specialist curators Metadata: specialized schema Reuse potential: high due to specialists Scope: discipline scope Discovery: known within discipline Specialist-Unaided Deposit: unaided deposit Metadata: specialized schema Reuse potential: moderate-high from discipline focus of metadata schema Scope: discipline scope Discovery: known within discipline Generalist-Unaided Deposit: unaided deposit Metadata: generalist schema Reuse potential: low as metadata is minimal Scope: discipline agnostic scope Discovery: broad name recognition i.e., institutional repositories
- 10. FEDERAL RESEARCH DATA SUMMARY • Observation networks and data portals are a fixed part of the landscape. They have a different role in open science than do repositories • Generalist repositories are easier to use than specialist repositories • Specialist repositories have higher reusability • Generalist repositories have economies of scale • If specialist repositories can leverage generalist repositories as back ends it would reduce overall cost
- 11. OPEN SCIENCE ROLE IN AI TRUSTWORTHINESS
- 12. “ON ARTIFICIAL INTELLIGENCE, TRUST IS A MUST, NOT A NICE-TO-HAVE” Margrethe Vestager, the European Commission executive vice president who oversees digital policy for the 27- nation bloc
- 13. TRUST ó TRUSTWORTHINESS TRUST • An individual’s confidence in an entity • “I trust this web site” TRUSTWORTHINESS • An entity’s state of being trustworthy or reliable • An estimate of an object’s worthiness to receive someone’s trust • Trustworthiness is difficult to accurately quantify
- 16. INDIANA UNIVERSITY BLOOMINGTON AI: Human-Machine Interaction § Fitness smartwatch, smart hearing aids § Co-bots, cyber-crews, digital twins § Integration of smart machines into human body in form of computer-brain interfaces or cyborgs AI: Autonomous and Semi- Autonomous Actors • Weapon systems • Robots in deep sea and space exploration • Self driving cars • Bots in financial trade AI: Big Data / Big Compute • Deep learning / Machine Learning / Natural Language Processing • Medical diagnosis, image recognition Broad Categories of AI
- 17. INDIANA UNIVERSITY BLOOMINGTON AI: Human-Machine Interaction § Fitness smartwatch, smart hearing aids § Co-bots, cyber-crews, digital twins § Integration of smart machines into human body in form of computer-brain interfaces or cyborgs AI: Autonomous and Semi- Autonomous Actors • Weapon systems • Robots in deep sea and space exploration • Self driving cars • Bots in financial trade AI: Big Data / Big Compute • Deep learning / Machine Learning / Natural Language Processing • Medical diagnosis, image recognition Broad Categories of AI Category with most urgency in issues of artificial moral agency
- 18. INDIANA UNIVERSITY BLOOMINGTON AI: Human-Machine Interaction § Fitness smartwatch, smart hearing aids § Co-bots, cyber-crews, digital twins § Integration of smart machines into human body in form of computer-brain interfaces or cyborgs AI: Autonomous and Semi- Autonomous Actors • Weapon systems • Robots in deep sea and space exploration • Self driving cars • Bots in financial trade AI: Big Data / Big Compute • Deep learning / Machine Learning / Natural Language Processing • Medical diagnosis, image recognition Broad Categories of AI Research needed in policy and technical extensions that lead to greater and more measurable forms of accountability
- 19. INDIANA UNIVERSITY BLOOMINGTON INTERVENTION POINTS: ENHANCED TRUSTWORTHINESS Developer ethics, development process norms Societal influence: public pressure, legislation, regulatory oversight AI algorithmic knowledge exhibiting higher levels of trustworthiness Technological manifestation: verifiable claims, explainability, accountability
- 20. Trustworthy AI is AI that is designed, developed, and used in a manner that is lawful, fair, unbiased, accurate, reliable, effective, safe, secure, resilient, understandable, and with processes in place to regularly monitor and evaluate the AI system’s performance and outcomes Lynne Parker, Deputy US Chief Technology Officer and Director of the National Artificial Intelligence Initiative Office
- 21. ML PROCESS M. Veale et al., CHI 2018 Data Training data Feature extraction Test data Learning algorithm Trained model Predict New data Explain- ability inquiries dev ops
- 22. RESEARCH PRODUCTS M. Veale et al., CHI 2018 Data Training data Feature extraction Test data Learning algorithm Trained model Predict New data Explain- ability inquiries dev ops
- 23. Open science contributes to trustworthy AI (trusted products) The research products of AI need to include intermediate results and explainability services