Representational Hierarchical Meta-Learning for Cross-Domain Few-Shot Learning

Abstract

Model-Agnostic Meta-Learning (MAML) is one of the most famous and basic algorithms in the meta- learning field. MAML devises learn-to-learn methods to hierarchically adapt a meta parameter to task- specific parameters for each task, where each task contains a few examples of a few classes. However, the diversity of tasks impacts the ability to generalize. That is to say, the meta-parameter that has been learned from the train tasks has to be generalized to the test tasks. When the test tasks are disjoint from the train tasks, the few adaptations from the meta-parameter using the test tasks are not sufficient to reach the optimal task-specific parameter. This kind of problem is addressed in the cross-domain few-shot-learning field and has been reviewed by various researchers. In this paper, a Representa- tional Hierarchical Meta-Learning is designed. In order to reinforce the adaptation ability of a meta- parameter, representational hierarchical meta-learning hierarchically tailors a global meta-parameter to domain-specific meta-parameters, and then adapts the domain-specific meta-parameters to task-specific parameters for the tasks in the corresponding domain. The experiments were carried out on several cross- domain few-shot-learning benchmarks, showing better but similar accuracies compared to the original MAML. The result implies that the representational hierarchical learning procedure was able to obtain a meta-parameter that better generalizes in the cross-domain setting. The inadequate performance gain was tried to explained with the derived generalization error bound.