Improving Generalization in Meta-Learning via Meta-Gradient Augmentation

Abstract

Meta-learning methods typically follow a two-loop framework, where each loop potentially suffers from notorious overfitting, hindering rapid adaptation and generalization to new tasks. Existing methods address this by enhancing the mutual-exclusivity or diversity of training samples, but these data manipulation strategies are data-dependent and insufficiently flexible. This work proposes a data-independent Meta-Gradient Augmentation (MGAug) method from the perspective of gradient regularization. The key idea is first to break the rote memories by network pruning to address memorization overfitting in the inner loop, then use the gradients of pruned sub-networks to augment meta-gradients, alleviating overfitting in the outer loop. Specifically, we explore three pruning strategies, including random width pruning, random parameter pruning, and a newly proposed catfish pruning that measures a Meta-Memorization Carrying Amount (MMCA) score for each parameter and prunes high-score ones to break rote memories. The proposed MGAug is theoretically guaranteed by the generalization bound from the PAC-Bayes framework. Extensive experiments on multiple few-shot learning benchmarks validate MGAug's effectiveness and significant improvement over various meta-baselines.