Learning to Balance: Bayesian Meta-Learning for Imbalanced and\n Out-of-distribution Tasks

Abstract

While tasks could come with varying the number of instances and classes in\nrealistic settings, the existing meta-learning approaches for few-shot\nclassification assume that the number of instances per task and class is fixed.\nDue to such restriction, they learn to equally utilize the meta-knowledge\nacross all the tasks, even when the number of instances per task and class\nlargely varies. Moreover, they do not consider distributional difference in\nunseen tasks, on which the meta-knowledge may have less usefulness depending on\nthe task relatedness. To overcome these limitations, we propose a novel\nmeta-learning model that adaptively balances the effect of the meta-learning\nand task-specific learning within each task. Through the learning of the\nbalancing variables, we can decide whether to obtain a solution by relying on\nthe meta-knowledge or task-specific learning. We formulate this objective into\na Bayesian inference framework and tackle it using variational inference. We\nvalidate our Bayesian Task-Adaptive Meta-Learning (Bayesian TAML) on multiple\nrealistic task- and class-imbalanced datasets, on which it significantly\noutperforms existing meta-learning approaches. Further ablation study confirms\nthe effectiveness of each balancing component and the Bayesian learning\nframework.\n