Neural Concept Learner: A Novel Approach to Learning Language

We need to learn both the mapping from the language to the program and the execution of the program at the same time. And so this is an extremely hard learning problem. You can think about it as like a spectrum, right? Like on one end of the spectrum, there's like fully end to end differentiable models. There's also work by Nikhish Gupta and Mike Lewis where kind of design your executor to be differentiable. So actually think this spectrum is super interesting in terms of how does the fact generalization affect sample complexity, how much data is needed? This is, I think, a really interesting direction.