Object Representations for Learning and Reasoning
Thirty-fourth Conference on Neural Information Processing Systems (NeurIPS)
December 11, 2020, Virtual Workshop
Understanding designed objects by program synthesis
- James McDermott
In this position paper, we argue that short programs in high-level, Turing-complete, human-readable symbolic languages are an attractive representation for designed objects: they have potential benefits both for general agents and for simple AI/ML tools. In principle they allow all available structure to be captured, and they allow for introspection. They may also parallel mechanisms of human thought. Actually finding these programs is a very difficult but worthwhile research problem.