Towards a measure of generalization ability.
To build machines that can program themselves to do any task, you need to measure how well they learn to do things they were never programmed to do.
Research
Fundamental work on program synthesis, toy environments, benchmarks and generative models.
