Attention Guided Relation Network for Few-Shot Image Classification
Imranul Ashrafi, Muntasir Mohammad, Arani Shawkat Mauree, Khan Mohammad Habibullah · 2019
Few-shot Learning is an object categorization problem where the classifier attempts to distinguish new classes with very few labeled examples. There has been significant progress in this field, which includes complex network architectures. Most of the works done in this field were focused on small datasets and longer training. In this paper, the experimentation was done with limited episodic training architecture, which consists of Relation Network as classification network, ResNet Embedding as embedding module, and Self Attention as attention mechanism. The experimentation and comparison with the state-of-the-art models show that attention with metric-based meta-learning generalizes quicker in short training and yields good results. The architecture was tested on the complex dataset miniImageNet. The accuracy was found to be 62.9%, which is close to the state-of-the-art architecture described on metric based meta-learning.