Subramanian is interested in both theoretical computer science and AI; having begun her graduate work in parallel program correctness and design, she found the questions of AI more intellectually appealing. She became tantalized, in her words, by ``the long-term scientific goal of understanding intelligence through computation and the shorter term goal of making machines smarter than they are today.''
In her doctoral research, Subramanian noted that today's systems, intelligent or otherwise, are limited by the conceptualization of the world given to them by their designers. When computational constraints on a task change, people have to reprogram the system. She constructed a theory of how an intelligent system faced with new computational pressures from the environment can automatically abstract its knowledge to solve its goals faster. This saves human designers from the tedium of reprogramming and allows systems to adapt to their environment by reprogramming themselves.
With one of her students, she has applied this theory to the task of optimizing functional programs. They have shown that the suite of optimizations discovered by people over the last 20 years can be automatically discovered by a machine in a few days.
Suggestions: