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

• DOI: 10.48550/arxiv.1905.12917
• Publication Date: 2019-01-01
• Authors: Donghyun Na; Haebeom Lee; Hayeon Lee; Saehoon Kim; Minseop Park; Eunho Yang; Sung Ju Hwang
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Statistics - Machine Learning; Machine Learning (stat.ML); 01 natural sciences; 0105 earth and related environmental sciences; Machine Learning (cs.LG)

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