Compositional Interfaces for Compositional Generalization

Jelena Luketina, Jack Lanchantin, Sainbayar Sukhbaatar, Arthur Szlam

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: compositional generalization, modular architectures, generalist agents

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We study how well can modular architectures generalize to unseen combinations of perceptual / action / instruction spaces

Abstract: With recent work such as GATO (Reed et al., 2022) we see the development of agents that can accomplish a variety of tasks, and are able to perceive the world and act in multiple observation and action spaces. We would want such agents to exhibit compositional generalization to unseen combinations of observation and action spaces, and adapt quickly to novel observation spaces by transfering knowledge. In this work, we demonstrate how these abilities can be achieved through the use of end-to-end modular architectures: the encoding of observations and the prediction of actions are handled by differentiable modules specialized to that space, with a single shared controller between them. To study the properties of such modular architectures in a controlled manner, we construct an environment with compositional structure, where each instance of the environment is created by combining an observation, action, and instruction space from a large set of options. We demonstrate that through the use of modularity, agents can generalize to unseen combinations of observation, action and instruction spaces; even when the unseen combinations are more challenging. Moreover, we demonstrate that modularity enables quick integration of novel observation modalities, requiring only adaptation of the modules encoding the new observation.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7644