Sanmi Koyejo, an assistant professor of computer science at Stanford University, is a rising star in the field of artificial intelligence (AI) and its applications in scientific discovery. His research focuses on the intersection of machine learning, AI, and the challenging question of trustworthiness in AI systems. With a background in electrical engineering and a passion for machine learning, Koyejo's journey to AI research was an unexpected one.
The Unexpected Path to AI
Koyejo's initial interest in electronics and control systems led him to explore communication technologies during his undergraduate studies. However, his path took an unexpected turn during his graduate school years. While studying cognitive radio systems, he discovered the power of machine learning as a tool to solve complex problems. This realization sparked his enthusiasm, and he quickly shifted his focus to machine learning, making it his primary research direction by the end of his PhD.
Expanding Horizons at Stanford
After completing his PhD, Koyejo joined Stanford University as a postdoc, where his interests expanded further. He became intrigued by the potential of AI to not only make predictions but also to assist scientists in making new discoveries and understanding the world better. This led him to establish the Stanford Trustworthy AI Research (STAIR) Lab, where his research group now works on three key areas:
- Understanding AI Systems: Exploring the capabilities and limitations of AI tools.
- Building Trustworthy AI: Enhancing the reliability, safety, and accuracy of AI systems.
- Applying AI to Real Problems: Utilizing AI in fields like healthcare, neuroscience, and astronomy.
Astronomy's Allure
Astronomy has become a significant area of interest for Koyejo's research group. He is drawn to the unique challenges and opportunities presented by astronomical problems. Unlike many machine learning success stories, astronomy often requires a blend of data, physical understanding, and scientific intuition. Koyejo finds this balance exciting, as AI can augment existing knowledge rather than replace it.
AI, Benchmarks, and Scientific Discovery
One of Koyejo's key messages is the distinction between performing well on a benchmark and doing science. While benchmarks are useful for comparing and tracking AI performance, they should not be misinterpreted as evidence of scientific achievement. Koyejo's research emphasizes the importance of evaluating AI systems in real-world, messy situations, rather than relying solely on controlled settings.
The Surprising Agreement of AI Systems
An intriguing finding from Koyejo's research is that AI systems often agree with each other, even when they are incorrect. This phenomenon raises questions about the assumption that agreement implies correctness. Koyejo argues that this agreement should prompt scientists to apply rigorous evidence standards, similar to those used in other scientific disciplines.
Scientists as AI Shapers
Koyejo also highlights the evolving role of scientists in the AI era. With the ease of producing research papers and the pressure on review systems, scientists should not view themselves as passive users of AI tools. Instead, they should actively participate in shaping the development and application of AI technologies, considering what counts as evidence and what makes a result trustworthy.
Mentorship and the Value of Deep Learning
In a world where technology enables rapid production, Koyejo emphasizes the importance of mentorship. He believes that good mentors help students develop judgment, perspective, and taste, which are crucial skills in an era where generating outputs is becoming easier. Koyejo encourages students to explore diverse fields and consider where they can make a meaningful impact, especially in fields like astronomy, which offer numerous open questions.
Looking Beyond the Hype
Koyejo's approach to AI research is balanced and evidence-driven. He avoids being overly optimistic or pessimistic, instead advocating for better questions and careful evaluation. His goal is to bridge the gap between impressions and headlines, moving towards a deeper understanding of AI's capabilities and limitations. This perspective may resonate with astronomers, who understand the importance of evidence and understanding in their field.
Join the Conversation
Sanmi Koyejo's plenary lecture at the AAS 248 meeting on June 15th promises to be a thought-provoking session. His insights into AI in astronomy and the evolving role of scientists in the AI landscape will undoubtedly spark interesting discussions. Stay tuned to learn more about Koyejo's work and the exciting possibilities at the intersection of AI and scientific discovery.
