Dual Meta Calibration Mix for Improving Generalization in Meta-Learning

Meta-learning has achieved remarkable success as a powerful paradigm for transferring learned knowledge from previous tasks. However, the lack of a large number of diverse and qualified tasks is the bottleneck of current meta-learning, which can easily lead to overfitting and therefore seriously hurt the generalization ability. In this paper, to address this challenge, we proposed Dual Meta Calibration Mix (DMCM) to improve the diversity and quality of tasks with "more data". Concretely, we designed a dual augmentation framework and meta calibration mix. The dual augmentation framework augments individual tasks and pairs of tasks by linearly combining samples and labels from both support and query sets, respectively. The meta calibration mix generates new samples by linearly combining image patches and corresponding labels based on the calibrated mixing matrix and calibrated label. Extensive experiments show that our proposed method significantly improves the generalization of meta-learning algorithms and consistently outperforms other state-of-the-art regularization meta-learning methods.