A Survey of Meta-learning for Classification Tasks

The superior performance of deep learning is supported by massive data and powerful computing engines. Meta-learning is an imitation of human learning ability. Instead of relying on massive quantities of data or numerous trials to learn features of current tasks, general knowledge obtained from historical tasks will be applied to future unknown tasks during meta-learning. Thus, it is considered one of the keys to achieving general artificial intelligence. In conjunction with the classification problem, meta-learning has had a new advancement recently, which is reviewed in this paper. First, the general settings and current formal definition of meta-learning are described. Then, the current methods in this field are summarized. The latest directions─methods based on data augmentation, transfer-learning, and unsupervised or semi-supervised learning are described in detail. Additionally, the quantitative performance of the examined approaches is assessed using benchmark datasets for categorization tasks facing small samples. Finally, it is proposed that the potential of meta-learning can be thoroughly explored from three perspectives: cross-domain adaptability, breakthrough of task space, and cost reduction.