Teaching Robots to Program

Long awaited technologies such as robot assistants or autonomous vehicles are far from reaching their potential to transform our society. Meanwhile, you have been hearing of the numerous deep learning success stories such as defeating the world champion in Go or recognising faces or playing ATARI games at a super human level. Why can’t we simply gather lots of human demonstrations of any task, crunch the data with a deep neural network and obtain state of the art performance?

Well, it turns out that our current approaches lack the capability of learning transparent and structured models from the endless amounts of data that we tend to throw at them. For example, explaining to someone how to play the simple ATARI game of Pong would not take you more than a couple of minutes, whereas state of the art algorithms need thousands of hours of gameplay. Abstracting models and relations from data is what we, as programmers, constantly do and programs provide the most general way of expressing such models. Therefore, in this talk we will have a look at the most general AI problem called inductive program synthesis and we will explore why and how we can teach robots to program.