2021 International Joint Conference on Neural Networks (IJCNN) | 2021
General meta-learning paradigm based on prior-models, meta-model, meta-algorithm, and few-shot-base-model
Abstract
As the studies on Machine Intelligence continue, new methods emerge with the purpose of improving computational learning, especially in deep-learning to overcome the need of high amount of data used to solve tasks. Therefore, paradigms like Meta Learning for few-shot learning intents to work out this situation. Nevertheless, current meta-learning methods address this situation using a single model trained in different levels of learning: meta-level and base-level. Instead, a new methodology to implement meta-learning for few-shot learning is proposed in this paper. This methodology is based on the concepts of modularity and reuse of optimized and previously implemented models and it is divided into elements of the process of learning. The elements are prior-models, meta-model, few-shot-base-model and a meta-algorithm. The proposed meta-algorithm generates a distribution over trained meta-models. Then, the best meta-model whose meta-tasks are more similar to the new tasks that the few-shot-base-model will solve is chosen. Also, a new concept of meta-few-shot sampling is proposed in order to evaluate how many meta-tasks are used in the meta-training stage. The experiments with 5-way 5-shot configuration show a competitive result compared with state-of-the-art methods. The proposed method achieves 97.56±1.57% and 83.33±8.33% in the CUB-200-2011 dataset, using the EfficientNetB7 and the DenseNet121 as prior-models, respectively; and, 99.56±0.63% (EfficientNetB7) and 94.89±4.28% (DenseNet121) in the Mini-ImageNet dataset.