Neurips Keynote 2020: Machine Learning as a Software Engineering Enterprise

The NeurIPS Keynote 2020: Machine Learning as a Software Engineering Enterprise presents a comprehensive discussion on the evolving role of Machine Learning (ML) as a discipline that must integrate principles from software engineering, programming languages, and social sciences to address complex technical and societal challenges.

Key Technological Concepts and Product Features:

• Machine Learning as Software Engineering:
  • ML is no longer just about algorithmic innovation or model accuracy; it requires rigorous software engineering practices.
  • Software engineering involves systematic methods, principles, and techniques to build robust, reliable, and maintainable systems.
  • ML systems must be treated as full software systems with attention to lifecycle, testing, deployment, and impact.
• Bias in Machine Learning:
  • Bias is pervasive and arises from multiple sources: data, algorithms, design decisions, and societal context.
  • Bias is not just a data problem but also a systems problem involving trade-offs and technical/social considerations.
  • Examples include facial recognition systems that perform poorly on people of color and gender-biased emotion recognition in healthcare robots.
  • Bias can be identified, quantified, and mitigated through specialized learning techniques, better data representation, and inclusive design.
• Algorithmic Fairness and Ethics:
  • Fairness definitions and privacy guarantees require precise formalization akin to theoretical computer science breakthroughs (e.g., public key cryptography).
  • Decisions about trade-offs (e.g., accuracy vs. fairness) must involve domain experts and affected communities, not just ML researchers.
  • Transparency and interpretability in high-stakes domains (criminal justice, healthcare, lending) are critical to prevent harm.
• Data and Model Robustness:
  • Models often exploit artifacts in training data and fail to generalize out-of-distribution.
  • Data augmentation is a practical but insufficient solution for robustness; deeper understanding and better abstractions are needed.
  • Increasing model complexity (more parameters) demands more data and can exacerbate bias and generalization challenges.
• Programming Languages and System Abstractions:
  • ML lacks well-defined abstractions and interfaces that enable composability, interpretability, and robustness.
  • Efforts to develop programming languages for ML (e.g., for reinforcement learning reward specification) aim to bridge gaps between domain experts and ML practitioners.
  • The idea of "sketching" tasks with partial program specifications that ML completes with data is proposed as a way to reason about bias and other properties.
• Diversity and Inclusion:
  • Diverse teams contribute to more robust, fair, and inclusive ML systems by bringing varied perspectives and identifying blind spots.
  • Underrepresentation of large population segments in computing and ML is a significant loss to the field and society.
  • Broadening participation should be institutionalized and valued in research and development environments.

Reviews, Guides, and Tutorials Provided:

• Historical and Social Context:
  • The keynote uses a creative narrative ("A Christmas Carol" style) to review ML research past, present, and future.
  • Historical examples highlight how physical device choices (e.g., camera optics, film) embed biases.
  • Present-day challenges with consumer internet data and individual-level predictions have intensified societal impacts.
• Case Studies:
  • Emotion recognition in healthcare robots showing gender bias and how specialized learners improved subgroup accuracy.
  • NLP models exploiting dataset artifacts leading to poor generalization.
  • StyleGAN and PULSE face generation systems exhibiting racial bias in upsampling images.
• Theoretical Foundations:
  • Emphasizes the importance of formal definitions for fairness, privacy, and robustness.
  • Draws parallels with the development of cryptography where rigorous definitions enabled progress.
  • Argues that theory is crucial for social impact-oriented ML research.
• Software Engineering Principles:
  • Distinguishes software engineering from mere coding.
  • Advocates for rigorous system design, ethical consideration, and lifecycle management in ML development.
  • Suggests that ML research must embrace these principles to ensure responsible deployment.
• Community and Collaboration:
  • Calls for multidisciplinary collaboration involving social scientists, domain experts, ethicists, and affected communities.
  • Highlights the need for accessible tools and interfaces to empower non-ML experts in decision-making.
  • Stresses listening to diverse voices and institutional changes to foster inclusivity.

Main Speakers and Sources:

• Charles Isbell – Key interlocutor and narrator, framing the discussion through a dialogue.
• Michael Littman – Represents a skeptical ML researcher initially focused on technical metrics.
• Ellie Pavlick – NLP researcher discussing dataset artifacts and generalization.
• Ayanna Howard – Robotics and human-robot interaction expert emphasizing inclusive design.
• Cynthia Rudin – Advocate for interpretable models and ethical ML applications.
• Michael Kearns – Theoretical computer scientist emphasizing the importance of formal definitions and interdisciplinary collaboration.
• Robert Ness – Researcher discussing racial bias in