Task Weighting in Meta-learning with Trajectory Optimisation

Cuong C. Nguyen, Thanh-Toan Do, Gustavo Carneiro

Abstract: Developing meta-learning algorithms that are un-biased toward a subset of training tasks often requires hand-designed criteria to weight tasks, potentially resulting in sub-optimal solutions. In this paper, we introduce a new principled and fully-automated task-weighting algorithm for meta-learning methods. By considering the weights of tasks within the same mini-batch as an action, and the meta-parameter of interest as the system state, we cast the task-weighting meta-learning problem to a trajectory optimisation and employ the iterative linear quadratic regulator to determine the optimal action or weights of tasks. We theoretically show that the proposed algorithm converges to an $\epsilon_{0}$-stationary point, and empirically demonstrate that the proposed approach out-performs common hand-engineering weighting methods in two few-shot learning benchmarks.

Submission Length: Long submission (more than 12 pages of main content)

Changes Since Last Submission:

We introduce the following major changes to address the main points identified by the reviewers:

• We split the Section Method into two subsections (Section 3.1 Task-weighting as a trajectory optimisation, and Section 3.2 Practical task-weighting method based on trajectory optimisation) for clarity.
• We move the Section Weight Visualisation from the Appendix to the main paper, making it a subsection of the Section Experiments.
• We add a new Section Ablation Studies, after the Section Experiments, to provide extensive analysis of the effect of some hyper-parameters used. In particular, we analyse the influence of the number of iLQR iterations, the length of the trajectory, the prior of the weighting vector $\mathbf{u}$ as well as training the three baselines more to have a fairer comparison, which takes more than 500 GPU-hour in total.
• We move the auxiliary lemmas in Section 4 to the Appendices to increase the readability of the Section Convergence Analysis.
• We add a brief description about the proposed method in the end of the Section Introduction to facilitate the understanding of our paper.
• We clarify further some prior studies in the Related Work (highlighted in Magenta) and move the Section Related Work to right before the Section Discussion and Conclusion.
• We provide an additional visualisation of shading plot requested by reviewer SZgw in Appendix J.

The changes above are all highlighted or annotated to facilitate the next round of reviews.

Code: https://github.com/cnguyen10/tow

Assigned Action Editor: Marcello Restelli

License: Creative Commons Attribution 4.0 International (CC BY 4.0)

Submission Number: 1145