No-regret Non-convex Online Meta-Learning

TL;DR

This paper extends online meta-learning to non-convex problems, introduces local regret as a performance measure, and demonstrates both theoretical guarantees and empirical superiority over traditional methods.

Contribution

It generalizes the online meta-learning framework from convex to non-convex settings and introduces local regret as a new performance metric.

Findings

Achieves logarithmic local regret in stochastic non-convex settings

Robust to hyperparameter initialization

Outperforms traditional methods on real-world tasks

Abstract

The online meta-learning framework is designed for the continual lifelong learning setting. It bridges two fields: meta-learning which tries to extract prior knowledge from past tasks for fast learning of future tasks, and online-learning which deals with the sequential setting where problems are revealed one by one. In this paper, we generalize the original framework from convex to non-convex setting, and introduce the local regret as the alternative performance measure. We then apply this framework to stochastic settings, and show theoretically that it enjoys a logarithmic local regret, and is robust to any hyperparameter initialization. The empirical test on a real-world task demonstrates its superiority compared with traditional methods.