Compositional generalization is a basic and essential intellective capability of human beings, which allows us to recombine known parts readily. However, existing neural network based models have been proven to be extremely deficient in such a capability. Inspired by work in cognition which argues compositionality can be captured by variable slots with symbolic functions, we present a refreshing view that connects a memory-augmented neural model with analytical expressions, to achieve compositional generalization. Our model consists of two cooperative neural modules, Composer and Solver, fitting well with the cognitive argument while being able to be trained in an end-to-end manner via a hierarchical reinforcement learning algorithm. Experiments on the well-known benchmark SCAN demonstrate that our model seizes a great ability of compositional generalization, solving all challenges addressed by previous works with 100% accuracies.

Additional Feedback: I think that the results on the few-shot MiniSCAN problem are very interesting and worth emphasizing more than they are emphasized currently. It's surprising and impressive that this approach can learn to generalize from such a limited dataset. I would definitely recommend finding more domains/problems which can show how well this approach works in the few-shot setting, because advances there would have very high value. Evaluation on more domains (particularly naturalistic ones) would also be helpful in showing that this approach can help solve compositional generalization problems in real-world settings. For instance, Nye et al (2020) show that the program synthesis approach can be applied to learn how to interpret number words in novel language from a limited number of examples; it would be really interesting to see if this approach could be applied to that problem or similar real-world problems.

Compositional generalization is a basic and essential intellective capability of human beings, which allows us to recombine known parts readily. However, existing neural network based models have been proven to be extremely deficient in such a capability. Inspired by work in cognition which argues compositionality can be captured by variable slots with symbolic functions, we present a refreshing view that connects a memory-augmented neural model with analytical expressions, to achieve compositional generalization. Our model consists of two cooperative neural modules, Composer and Solver, fitting well with the cognitive argument while being able to be trained in an end-to-end manner via a hierarchical reinforcement learning algorithm. Experiments on the well-known benchmark SCAN demonstrate that our model seizes a great ability of compositional generalization, solving all challenges addressed by previous works with 100% accuracies.