Ai for the rest of us: Digital Discrimination: Cognitive Bias in Machine Learning (and LLMs!)

Resources from this talk:

Paper: “Foundation models: Opportunities, risks and mitigations” https://www.ibm.com/downloads/cas/E5KE5KRZ

Racist/Sexist AI Generated Imagery
https://arxiv.org/pdf/2110.01963
https://www.businessinsider.com/lensa-ai-raises-serious-concerns-sexualization-art-theft-data-2023-1
https://www.technologyreview.com/2022/12/12/1064751/the-viral-ai-avatar-app-lensa-undressed-me-without-my-consent/

Healthcare Inequality
https://www.forbes.com/sites/adigaskell/2022/12/02/minority-patients-often-left-behind-by-health-ai/?sh=31d28a225b41
https://medicine.yale.edu/news-article/eliminating-racial-bias-in-health-care-ai-expert-panel-offers-guidelines/
https://www.nejm.org/doi/full/10.1056/NEJMms2004740
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11228769/

Gender Shades/Algorithmic Justice League
http://gendershades.org/
https://www.youtube.com/watch?v=Af2VmR-iGkY
https://www.ajl.org/take-action
https://www.netflix.com/title/81328723

Education
https://www.nytimes.com/2018/02/12/business/computer-science-ethics-courses.html
https://d4bl.org/

National and Industry Standards
https://lists.lfaidata.foundation/g/gac-responsible-ai-workstream
https://genaicommons.org/
Responsible Ai Workstream on Slack: https://ibm.biz/responsible-ai-workstream-slack
https://www.nist.gov/aisi
https://ec.europa.eu/digital-single-market/en/news/ethics-guidelines-trustworthy-ai
https://artificialintelligenceact.eu/
https://ai.gov/
https://digitalservices.vermont.gov/ai

Tools to combat AI Bias
https://landscape.lfai.foundation/card-mode?category=trusted-responsible-ai&grouping=category
https://huggingface.co/ibm-granite
https://newsroom.ibm.com/2024-10-21-ibm-introduces-granite-3-0-high-performing-ai-models-built-for-business
https://instructlab.ai/

Algorithmic Justice League
https://www.ajl.org/
https://secure.actblue.com/donate/algorithmic-justice-league

Transcript

1. Digital Discrimination: Cognitive Bias in Machine Learning (and LLMs!) Mo McElaney Open Source Developer Programs at IBM AI for the rest of us 1
2. About me… 2
3. About me… 3
4. Open Source AI @ IBM 4
5. SLIDES! Digital Discrimination: Cognitive Bias in Machine Learning (and LLMs!) 5 https://bit.ly/digital-discrimination-london
6. Agenda • Examples of Bias in Machine Learning. • Solutions to combat unwanted bias. • Tools to combat unwanted bias. • Resources and how to get involved. 6
7. A cognitive bias is a systematic pattern of deviation from norm or rationality in judgment. People make decisions given their limited resources. Wilke A. and Mata R. (2012) “Cognitive Bias”, Clarkson University 7
8. Examples of bias in machine learning… 8
9. Racist and sexist image generation January 2023 Image Credit: #WOCinTech 9
10. https://www.technologyreview.com/2022/12/12/1064751/the-viral-ai-avatar-app-lensa-undressed-me-without-my-consent/ 10
11. “In 1,000 years, when we look back as we are generating the thumbprint of our society and culture right now through these images, is this how we want to see women?” - Melissa Heikkilä Senior reporter at MIT Technology Review, covering artificial intelligence and how it is changing our society. 11
12. Healthcare Inequality Dec 2022 Image Credit: #WOCinTech 12
13. Machine Learning Pipeline In- Processing Pre- Processing Post- Processing Modifying the training data. Modifying the learning algorithm. Modifying the predictions (or outcomes.) 13
14. Yale Panel’s Guidelines on Eliminating Racial Bias in Health Care AI Mitigating algorithmic bias, must take place across all stages of an algorithm’s life cycle. The experts defined this life cycle in five stages: 1. Identification of the problem that the algorithm will address 2. Selection and management of data to be used by the algorithm 3. Development, training, and validation of the algorithm 4. Deployment of the algorithm 5. Ongoing evaluation of performance and outcomes of the algorithm 14
15. Yale Panel’s Guidelines on Eliminating Racial Bias in Health Care AI Five guiding principles for preventing algorithmic bias: 1. Promote health and health care equity during all phases of the health care algorithm life cycle 2. Ensure that health care algorithms and their use are transparent and explainable 3. Authentically engage patients and communities during all phases of the health care algorithm life cycle and earn trust 4. Explicitly identify health care algorithmic fairness issues and tradeoffs 5. Ensure accountability for equity and fairness in outcomes from health care algorithms 15
16. “As the use of new AI techniques grows and grows, it will be important to watch out for these biases to make sure we do no harm to specific groups while advancing health for others. We need to develop strategies for AI to advance health for all.” Lucila Ohno-Machado, MD, PhD, MBA
17. Gender Shades Project February 2018 Image Credit: #WOCinTech 17
18. http://gendershades.org/ 18
19. “If we fail to make ethical and inclusive artificial intelligence we risk losing gains made in civil rights and gender equity under the guise of machine neutrality.” - Joy Boulamwini @jovialjoy
20. http://www.aies-conference.com/wp-content/uploads/2019/01/AIES-19_paper_223.pdf 20
21. https://www.youtube.com/watch?v=Af2VmR-iGkY 21
22. Agenda • Examples of Bias in Machine Learning. • Solutions to combat unwanted bias. • Tools to combat unwanted bias. • Resources and how to get involved. 22
23. Solutions? What can we do to combat bias in AI? 23
24. EDUCATION IS KEY Image Credit: #WOCinTech 24
25. https://www.nytimes.com/2018/02/12/business/computer-science- ethics-courses.html 25
26. Questions posed to students in these courses... Is the technology fair? How do you make sure that the data is not biased? Should machines be judging humans? 26
27. https://twitter.com/Neurosarda/status/1084198368526680064 27
28. National and Industry Standards Image Credit: #WOCinTech 28
29. https://lists.lfaidata.foundation/g/gac-responsible-ai-workstream 29
30. Responsible AI Workstream at the Generative AI Commons The Responsible AI Framework We are a community of industry professionals, students, academics, experts, practitioners, and enthusiasts. We meet every other Thursday at 4pm Central Europe. Your contributions are most welcome. Join us on Slack Draft made public at Ai_dev Hong Kong In August, 2024 https://genaicommons.org/ 30 Find us in the #gac-responsible-ai-workstream channel!
31. NIST AI Safety Institute Consortium (AISIC) The Consortium brings together more than 200 organizations to develop science-based and empirically backed guidelines and standards for AI measurement and policy. This will help ready the U.S. to address the capabilities of the next generation of AI models or systems, from frontier models to new applications and approaches, with appropriate risk management strategies. LF AI and Data is active in NIST AISIC. Workgroups: •wg #1 - Risk Management for Generative AI •wg #2 - Synthetic Data •wg #3 - Capability Evaluations •wg #4 - Red Teaming •wg #5 - Safety & Security https://www.nist.gov/aisi 31
32. AI Regulations and Guidelines are emerging in the EU and the US 32
33. EU Ethics Guidelines for Trustworthy Artificial Intelligence According to the Guidelines, trustworthy AI should be: (1) lawful - respecting all applicable laws and regulations (2) ethical - respecting ethical principles and values (3) robust - both from a technical perspective while taking into account its social environment Source: https://ec.europa.eu/digital-single-market/en/news/ethics-guidelines-trustworthy-ai 33
34. EU Artificial Intelligence Act #1 - Human agency and oversight. #2 - Technical robustness and safety. #3 - Privacy and data governance. #4 - Transparency #5 - Diversity, non- discrimination and fairness. #6 - Societal and environmental well-being. #7 - Accountability US Executive Order on AI #1 - Ensure responsible and effective government use of AI #2 - Ensure safety and security #3 - Protect Americans’ privacy #4 - Transparency #5 - Advance equity and civil rights #6 - Stand up for consumers and workers #7 - Accountability Different: Promote innovation and competition Advance American leadership abroad 34
35. AI Regulations in the UK 35
36. UK’s Treaty Governing Safe Use of Ai The new agreement has 3 over-arching safeguards: • protecting human rights, including ensuring people’s data is used appropriately, their privacy is respected and AI does not discriminate against them • protecting democracy by ensuring countries take steps to prevent public institutions and processes being undermined • protecting the rule of law, by putting the onus on signatory countries to regulate AI-specific risks, protect its citizens from potential harms and ensure it is used safely Source: https://www.gov.uk/government/news/uk-signs-first-international-treaty-addressing-risks-of-artificial-intelligence 36
37. Agenda • Examples of Bias in Machine Learning. • Solutions to combat unwanted bias. • Tools to combat unwanted bias. • Resources and how to get involved. 37
38. TOOLS TO COMBAT BIAS Image Credit: #WOCinTech 38
39. Tool #1: Trusted AI Projects at LF AI & Data Open Source Libraries 39
40. Trusted AI projects at LF AI & Data https://landscape.lfai.foundation/card-mode?category=trusted-responsible-ai&grouping=category 40
41. Adversarial Robustness 360 ↳ (ART) AI Fairness 360 ↳ (AIF360) AI Explainability 360 ↳ (AIX360) https://github.com/Truste d-AI/AIF360 FAIRNESS EXPLAINABILITY ROBUSTNESS Trusted AI Lifecycle through Open Source Pillars of trust, woven into the lifecycle of an AI application https://github.com/Truste d-AI/adversarial-robustne ss-toolbox https://github.com/Trust ed-AI/AIX360 Is it fair? Is it easy to understand? Did anyone tamper with it? 41
42. Tool #2: Granite Open Source Large Language Models 42
43. IBM Granite 61% of CEOs identify concerns about data lineage and provenance as a barrier to adopting generative AI. • Intellectual property (IP) indemnity protection • Built for the enterprise • Built to minimize hateful and profane content, or “HAP.” • Strong focus on data governance, risk and compliance • Open Source and available now on Hugging Face! • https://huggingface.co/ibm-granite 43
44. Granite 3.0!!! https://www.ibm.com/granite My fav highlights: • Disclosure of the enterprise relevant datasets used • Process for validating the license associated with all these datasets, Granite is open source under an Apache 2.0 license! • Models are trained on 12T tokens, that is 3-6x more data than any Granite trained before and new advanced capabilities like tool calling. • Developer enablement - recipes plus a playground and documentation, plus partnerships with ollama, replicate, nvidia, etc • And more! 44
45. Tool #3: InstructLab Open Source-style contributions to LLMs 45
46. InstructLab: A community-powered approach for building truly Open Source LLMs 46
47. The Problems with LLMs 47 •There is no path to merge and combine open source contributions to LLMs •Instead of contributing back, users create new variants of the LLM •The community is not able to share with each other all the model improvements •Contributions built on Llama result in… more llamas… •What we need is an approach that allows the whole community to benefit from model improvement contributions
48. What is InstructLab? 48 • InstructLab is an open source project for AI that welcomes everyone to help improve LLMs • It allows people to easily contribute updates and changes to these AI models • It makes it simple for anyone to get involved and shape the future of generative AI enterprise • It allows creators of AI models to easily update and improve their models without having to start from scratch. • Instead of rebuilding and retraining the whole model, simply add new skills and knowledge to it https://github.com/instructlab The model stack The community can create and contribute knowledge & skills recipes
49. Summary of Solutions Transparent • Tools for Explainability • Open Source AI Examples: ◦ Trusted and Responsible AI projects at LF AI ◦ Granite ◦ InstructLab Proactive Accountable • Tools addressing Bias and Fairness • Industry Standards • Government Legislation 49 • Education Requirements • Testing for Adversarial Robustness • Responsible AI Workstreams • Treaties establishing shared language
50. Agenda • Examples of Bias in Machine Learning. • Solutions to combat unwanted bias. • Tools to combat unwanted bias. • Resources and how to get involved. 50
51. https://www.ajl.org/ 51
52. https://secure.actblue.com/donate/algorithmic-justice-league 52
53. Photo by rawpixel on Unsplash No matter what it is your responsibility to build systems that are fair. 53
54. Thank You! Find Mo McElaney on LinkedIn: 54
55. Resources from this talk: Paper: “Foundation models: Opportunities, risks and mitigations” https://www.ibm.com/downloads/cas/E5KE5KRZ Racist/Sexist AI Generated Imagery https://arxiv.org/pdf/2110.01963 https://www.businessinsider.com/lensa-ai-raises-serious-co ncerns-sexualization-art-theft-data-2023-1 https://www.technologyreview.com/2022/12/12/1064751/t he-viral-ai-avatar-app-lensa-undressed-me-without-my-cons ent/ Healthcare Inequality https://www.forbes.com/sites/adigaskell/2022/12/02/minori ty-patients-often-left-behind-by-health-ai/?sh=31d28a225b 41 https://medicine.yale.edu/news-article/eliminating-racial-bia s-in-health-care-ai-expert-panel-offers-guidelines/ https://www.nejm.org/doi/full/10.1056/NEJMms2004740 https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11228769/ Gender Shades/Algorithmic Justice League http://gendershades.org/ https://www.youtube.com/watch?v=Af2VmR-iGkY https://www.ajl.org/take-action https://www.netflix.com/title/81328723 Education https://www.nytimes.com/2018/02/12/business/computer-scien ce-ethics-courses.html https://d4bl.org/ National and Industry Standards https://lists.lfaidata.foundation/g/gac-responsible-ai-workstrea m https://genaicommons.org/ https://www.nist.gov/aisi https://ec.europa.eu/digital-single-market/en/news/ethics-guide lines-trustworthy-ai https://artificialintelligenceact.eu/ https://ai.gov/ https://digitalservices.vermont.gov/ai Tools to combat AI Bias https://landscape.lfai.foundation/card-mode?category=trusted-r esponsible-ai&grouping=category https://huggingface.co/ibm-granite https://newsroom.ibm.com/2024-10-21-ibm-introduces-granite-3-0-hi gh-performing-ai-models-built-for-business https://instructlab.ai/ Algorithmic Justice League https://www.ajl.org/ https://secure.actblue.com/donate/algorithmic-justice-league 55 https://bit.ly/digital-discrimination-london