Sanmi Koyejo, an assistant professor of computer science at Stanford University, is a leading figure in the field of trustworthy AI research. His work focuses on the intersection of machine learning, scientific discovery, and the complex question of how to trust AI systems. In this interview, Koyejo discusses his journey into AI, his research interests, and his passion for astronomy.
Koyejo's interest in AI began unexpectedly during his graduate studies. Initially, he was more inclined towards electrical engineering, building electronics, and working on control systems. However, his work on cognitive radio systems, which required the use of machine learning tools, sparked his excitement. This led him to pursue machine learning as his primary research direction during his PhD.
After completing his PhD, Koyejo joined Stanford University for a postdoc, where he became interested in the broader implications of AI. He wanted to explore not only the predictive capabilities of AI but also its potential to assist scientists in making new discoveries and understanding the world better. Today, his research group works in three main areas: understanding AI systems, building trustworthy AI, and applying AI to real-world problems, particularly in science and healthcare.
One of Koyejo's key research interests is in astronomy. He finds the problems astronomers work on particularly intriguing because they differ from the typical scenarios where AI excels. Astronomy often requires combining observations with physical understanding and scientific intuition, as there is only one universe to work with. Koyejo's early work in astronomy has involved collaborating with researchers to explore how machine learning can enhance the efficiency of analysis methods while preserving the underlying physical understanding.
Koyejo emphasizes the importance of distinguishing between doing well on a test and doing science. AI researchers often evaluate systems using benchmarks, which are standard collections of problems designed to measure performance. While these benchmarks are useful for comparison and tracking progress, Koyejo warns against overinterpreting the results. He argues that passing a benchmark does not necessarily mean the AI system can solve larger problems. His research aims to develop better evaluation methods, focusing on whether AI systems can work in real-world, messy situations.
One surprising finding from Koyejo's research is that AI systems often agree with each other, even when they are incorrect. This leads him to emphasize the need for applying the same standards of evidence that scientists use in other fields. Koyejo believes that scientists should actively shape the development of AI tools, as they understand what constitutes evidence, which mistakes matter, and what makes a result trustworthy.
In terms of advice for students, Koyejo stresses the importance of mentorship. With technology making it easier to produce work quickly, understanding the work becomes even more crucial. Good mentors help students develop judgment, perspective, and taste, which are valuable skills in an era where generating outputs is easier than ever. Koyejo encourages students to explore widely and think deeply about their impact, as fields like astronomy offer exciting opportunities for scientific discovery.
Overall, Koyejo's work highlights the need for a careful and nuanced approach to AI research. He sees a real excitement in the potential of AI systems but also emphasizes the importance of careful evaluation before making big claims. By focusing on closing the gap between impressions and headlines and evidence and understanding, Koyejo's research contributes to the development of trustworthy AI in various fields, including astronomy.
