seminar

This week, we covered a wide variety of approaches to neural methods to program synthesis and program induction. The first batch of presentations: Romila, Eric and Henri Romila: we have synthesis; the most naive way is indeed one way that CAN be used to form programs: get the input space, the program space, the action space, and then try all possible combinations. Her paper talked about some sort of Monte Carlo tree search based method; alpha zero but applied to program synthesis, and she was critical as they don’t update the policy network every time. [Read More]