Wednesday / CS Seminar

February 1st / 12:00 PM

Hybrid / Mudd 3514

Title: Interpreting AI for Science Models: Why Explainability, Scrutiny, and Transparency Enable Scientific Discovery, Not Limit It

Speaker: Austin Clyde

Zoom Link

Livestream

Abstract:

Despite artificial intelligence's (AI) achievements in numerous scientific challenges, many fear that AI threatens human creativity and intelligence. These fears echo an ancient myth that writing would weaken memory and lead to only a semblance of understanding. Unlike this myth's fears, today, we find the interpretation of texts as a source of new possibilities and freedoms, not deferent regurgitation. Interpretation is not just for text but enables a range of freedoms, from developing new human rights to scientific understanding. In this talk, I argue that interpretability needs to be a first-class principle in data science education, engineering, and theory--not a post-hoc consideration--to meet global citizens' most pressing needs from science. Using examples from my research, I show how taking interpretation as a core concern leads to state-of-the-art technical solutions, such as accelerating early-stage drug discovery platforms 100-fold and developing novel solutions to the chemical enumeration problem with large-language models. Motivated by these examples, I will discuss how this conceptual shift reveals new paths for explainability research, such as treating the model development process as causal. Explainability in this plural way has exciting implications for numerous legal and human rights challenges in deploying automated decision systems. By opening data science to interpretative flexibility and providing a humanistic education within data science, citizens can contest, argue, and shape claims made by and about AI decision-making systems. This core component of agency, the freedom to interpret and shape claims, is critical to scientific success and fostering well-functioning democracies.

Biography:

Austin Clyde is an Assistant Computational Scientist in the Data Science and Learning Division at Argonne National Laboratory. He holds a Ph.D. in computer science from the University of Chicago, where he continues to lecture at the Pozen Family Center for Human Rights. His research has received the ACM Gordon Bell Special Prize for HPC-Based COVID-19 Research in 2020 and 2022 and a Department of Energy Secretary's Honor Award for involvement in the National Virtual Biotechnology Laboratory's COVID-19 Response. Previously, he was a visiting research fellow at the Harvard Kennedy School's Program on Science, Technology, and Society.