IHPST Research Seminar - First Session - Winter 2025

You are cordially invited to attend the Winter 2025 Research Seminar Series of the Institute for the History and Philosophy of Science and Technology, organized by Assistant Professor Karina Vold

“Artificial intelligence benchmarks and degenerating research”

Hybrid event

Abstract:

The move toward creating massive benchmark datasets, starting with ImageNet, ushered in the rise to prominence of deep learning, and along with it the proliferation of benchmarks for evaluating AI models at performing tasks like recognizing faces, translating languages, and captioning images. Although benchmarks have undeniably been key tools in the rise of deep learning, some researchers have critiqued an overly enthusiastic embrace of benchmarks (a.k.a. benchmark chasing) as a harmful practice, both ethically and epistemically.

Benchmarks can reify narrow views of correct task performance, distort research incentives, and favor researchers at wealthy institutions that can most afford to train state-of-the-art models. Recent empirical evidence has begun to track and quantify the effects and magnitude of the bias toward industry-sponsored versus academic research in areas like natural language processing, raising concerns about corporate capture of research.

Criticism of benchmark chasing is often met with pushback. Defenders of AI insist that critics are trying to move the goalposts each time models manage to surpass a benchmark, and the debate has sometimes devolved to name calling. Drawing on the literature on scientific progress, I argue that critique should be embraced as an essential component in a progressive research program, rather than silenced, and that the boundaries around who is considered a relevant expert worth heeding are too narrow.

Catherine Stinson is Queen’s National Scholar in the Philosophical Implications of Artificial Intelligence and assistant professor in computing and philosophy at Queen’s University. Stinson’s previous positions include postdocs at the Max Planck Institute for Biological Cybernetics and the Leverhulme Centre for the Future of Intelligence at the University of Cambridge, and senior policy associate at the Mowat Centre, University of Toronto. With graduate training in both machine learning and philosophy of science, plus experience in policy and visual art, Stinson brings multiple perspectives to the study of AI. The research in their Ethics & Technology Lab includes methodological analyses of AI research, data-driven sociological studies of AI as a field, empirical work in ethical AI, technical support for data justice initiatives, and collaborations with artists.

Wednesday, February 12, 2025, 12:30- 2:00 p.m. Rotman School of Management, 95 St, George Street, Room 1065

Seminar will be broadcast live via Zoom (register for link)

This seminar is co-hosted by the University of Toronto’s Schwartz Reisman Institute for Technology and Society, and will be presented in-person with an online option at the University of Toronto’s Rotman School of Management.