Keller Colloquium in Computing and Mathematical Sciences

Program synthesis traditionally has dealt with the problem of automatically generating programs from very high-level descriptions of their behavior. In the early days of the field, it was assumed that these descriptions would take the form of complete formal descriptions of the behavior, but more recently, this requirement has been relaxed to support weaker specifications such as test harnesses or input/output examples. This more general view of the synthesis problem has blurred the line between program synthesis and Machine learning, creating new opportunities both to apply Machine learning techniques to attack the kind of automatic programming problems that program synthesis has traditionally tackled, as well as to apply program synthesis techniques to attack Machine Learning problems. In this talk I will describe some recent work by my group exploring this boundary between program synthesis and Machine learning. I will describe how program synthesis techniques rooted in formal methods can be combined with machine learning techniques can help us provide new solutions to interesting problems including classification, modeling and control, in addition to automated programming.