Analysis Regarding The Learning-To-Learn Process In The Implementation Of A Meta-Supervised Algorithm For Few-Shot Learning

Currently, in the efforts of consolidating models based on theories that simulate cognitive and learning human processes, the Meta-Learning paradigm (MeL) has emerged in the Machine Intelligence area. MeL has been applied in combination with other approaches for Machine Learning and Deep-Learning, for instance, to solve new tasks with few samples in the few-shot learning approach. However, to consider MeL as a well-defined learning paradigm aimed at successfully simulating learning human processes, it is essential to find a from to evaluate its learning-to-learn process (LTLP). Metrics reported in the literature to evaluate the LTLP are insufficient and shallow to understand the results reported by MeL-based models. The most common metric used to evaluate and compare MeL-based models is the metric accuracy. Therefore, this work proposes the use of the Centered Kernel Alignment (CKA) metric to evaluate the LTLP of MeL-based models for few-shot learning. The CKA metric helps to quantify how much of the meta-knowledge and meta-experience learned in the meta-training of the meta-models are used to solve new tasks, contributing to the understanding of the LTLP. A MeL methodology is implemented using 9 prior-models in the meta-training of meta-models. In the experiments of the meta-models, the CKA metric is used to analyze similarities between the meta-features obtained by each prior-model. Besides, a proposed meta-supervised algorithm uses the CKA metric to select the meta-parameters to solve new tasks in the base-domain. The results showed CKA scores of 0.674 and 0.799 (these scores mean the level in which the new base-tasks to solve are different from the meta-learned tasks), and accuracies of 92.27% and 99.73%, for solving the new base-tasks using the CUB and Mini-ImageNet datasets, respectively.