Mo McElaney- Digital Discrimination: Cognitive Bias in Machine Learning (and LLMs!) | PyData Vermont

Ai

Explore cognitive bias in machine learning and LLMs with Mo McElaney. Learn about real-world examples of AI discrimination, current solutions, and strategies for developing more equitable AI systems.

Key takeaways
  • Cognitive bias in machine learning reflects human biases, as AI systems are trained on data created by humans and algorithms designed by humans

  • Notable examples of AI bias include:

    • Lensa app generating hypersexualized and racialized images, particularly affecting Asian women
    • Healthcare algorithms perpetuating systemic discrimination through biased training data
    • Gender Shades Project revealing significant accuracy gaps in facial recognition systems across gender and skin color
  • Key solutions for combating AI bias:

    • Implementing “train then mask” approaches to handle sensitive data
    • Requiring computer science students to take ethics courses
    • Using open source tools for transparency and accountability
    • Following established guidelines and frameworks (EU AI Act, US AI Safety Institute)
  • Organizations working on AI bias include:

    • Linux Foundation AI (LFAI)
    • National Institute for Standards and Technology (NIST)
    • Various university research programs
    • IBM’s Granite project
  • Important principles for responsible AI development:

    • Transparency in model training and data sources
    • Accountability for outcomes
    • Regular evaluation and monitoring
    • Community engagement
    • Protection of privacy and individual rights
  • Regulatory landscape:

    • Over 37 countries have proposed AI regulation frameworks
    • EU focuses on ethics, transparency, and accountability
    • US emphasizes innovation and competitive leadership
    • Vermont implementing state-level AI guidelines and ethics
  • Bias mitigation requires intervention at all stages:

    • Data collection and selection
    • Algorithm development
    • Training and validation
    • Deployment
    • Ongoing monitoring and adjustment