Advancements in Meta-Learning Paradigms: A Comprehensive Exploration of Techniques for Few-Shot Learning in Computer Vision

The rapid evolution of computer vision applications demands innovative approaches to address the challenges posed by limited labelled data. This study undertakes a comprehensive exploration of meta-learning paradigms, specifically focusing on their efficacy in the realm of few-shot learning. Three distinct meta-learning frameworks (MAML, Reptile and Matching networks) are scrutinised and compared to elucidate their contributions to the adaptability and generalisation capabilities of computer vision models. This research contributes valuable insights to the meta-learning discourse in computer vision, offering nuanced understandings of framework strengths and limitations and serving as a catalyst for future innovations in adaptive machine learning models. This study also compares performance of MAML, Reptile and their variants and discusses applications of FSL in computer vision. This study concludes with nuanced differences in performance with higher F1 score and accuracy in Reptile framework, having a better trade-off among recall and precision.