Meta-Learning with Explicit Task Information

Gabriel Tseng, Hannah Kerner, David Rolnick

Keywords: meta-learning, climate change, agriculture, remote sensing

TL;DR: A meta-learning algorithm which incorporates task-specific metadata to learn context across tasks

Abstract: A common approach in few-shot learning is to adapt to a new task after learning a variety of similar tasks. When the diversity of the tasks is high, however, it can be challenging for models to generalize effectively. Prior work has approached this problem by inferring task information implicitly from the data in order to better adapt to each new task. However, in some cases, explicit information about tasks is available that can inform task adaptation to improve performance, especially in the context of few-shot learning. In this work, we introduce task-informed meta-learning (TIML), an algorithm which modulates a model based on explicit task metadata. We evaluated TIML for a range of classification and regression tasks and found that TIML significantly improves performance in both regimes across a diversity of model architectures. In particular, we show the power of TIML in remote sensing for agriculture---an area of high societal impact where traditional methods have failed due to limited and imbalanced data.

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.

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

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)