On the usefulness of decomposition and composition strategies in program synthesis

Abstract: Problem decomposition - decomposing a complex problem into a simpler sub-problems - is a successful strategy to tackle complex problems. In this talk, I will show how pursuing this idea helps us to develop better AI systems, focusing on program synthesis. Namely, I will show how decomposing programs into subprograms helps to more efficiently solve new synthesis  tasks, how changing the representation of data by composing be relations helps to better capture generative processes, and finally how can we design effective hybrid neuro-symbolic systems by decomposing the functional roles in such a system.