[edit]
Challenges of Acquiring Compositional Inductive Biases via Meta-Learning
AAAI Workshop on Meta-Learning and MetaDL Challenge, PMLR 140:138-148, 2021.
Abstract
Meta-learning is typically applied to settings where, given a distribution over related training tasks, the goal is to learn inductive biases that aid in generalization to new tasks from this distribution. Alternatively, we might consider a scenario where, given an inductive bias, we must construct a family of tasks that will inject the given inductive bias into a parametric model (e.g. a neural network) if meta-training is performed on the constructed task family. Inspired by recent work showing that such an algorithm can leverage meta-learning to improve generalization on a single-task learning problem, we consider various approaches to both a) the construction of the family of tasks and b) the procedure for selecting support sets for a particular single-task problem, the SCAN compositional generalization benchmark. We perform ablation experiments aimed at identifying when a meta-learning algorithm and family of tasks can impart the compositional inductive bias needed to solve SCAN. We conclude that existing meta-learning approaches to injecting compositional inductive biases are brittle and difficult to interpret, showing high sensitivity to both the family of meta-training tasks and the procedure for selecting support sets.