A Systematic Literature Review of Meta-Learning Models for Classification Tasks

Abstract

Meta-learning is a field of learning that aims at addressing the challenges of conventional machine learning approaches such as learning from scratch for every new task. The main aim of this study was to do a systematic literature review of the existing meta-learning models that have been developed, published, and can be used for classification tasks. Systematic literature review method was used, employing a search of journal articles and publications of conference proceedings. The process involved data collection, analysis, and reporting of the results. To achieve the objective, 30 primary papers published since 2016 and relevant to classification tasks in meta-learning were considered. Data was extracted from the papers, then the following was analyzed in each model as presented in the papers; techniques used, the contribution, and the research gap. Although a lot has been done so far in Meta-learning, the existing models are not yet optimal. They still have challenges in few-shot learning, computation time complexity, difficulty in continual learning, and generalizability across multiple related tasks during transfer learning.