Multi-Initialization Meta-Learning with Domain Adaptation

Abstract

Recently, meta learning providing multiple initializations has drawn much attention due to its capability of handling multi-modal tasks drawn from diverse distributions. However, because of the difference of class distribution between meta-training and meta-test domain, the domain shift occurs in multi-modal meta-learning setting. To improve the performance on multi-modal tasks, we propose multi-initialization meta-learning with domain adaptation (MIML-DA) to tackle such domain shift. MIML-DA consists of a modulation network and a novel meta separation network (MSN), where the modulation network is to encode tasks into common and private modulation vectors, and then MSN uses these vectors separately to update the cross-domain meta-learner via a double-gradient descent process. In addition, the regularization using inequality measure is considered to improve the generalization ability of the meta-learner. Extensive experiments demonstrate the effectiveness of our MIML-DA method to new multi-modal tasks.