Deep Learning with Limited Data: Advanced Classification Approaches Through Few-Shot Learning and Prototype Networks

Classification problems in the fields of machine learning and artificial intelligence facilitate the extraction of meaningful information from data by assigning inputs to specific categories. Classification processes offer solutions for a wide range of areas, including health, agriculture, education, and sports. However, the classification process typically requires a large amount of labeled data. Accessing a large volume of labeled data is costly and time-consuming. The few-shot learning method has been utilized to address this issue, allowing models to learn new tasks with minimal examples. In this article, pre-trained deep network architectures have been fed into prototype networks, creating representative examples for each class. Thus, the category to which new data belongs is determined based on its similarity to the prototypes. Experimental studies have been conducted on the Food101 and Oxford-III Pet datasets, and the experimental results have been measured using four different evaluation metrics. The results have been presented and interpreted both in table form and graphically. In comparing classification accuracy, the metrics of Accuracy, F1_Score, Precision, and Recall were utilized. For the Oxford-III Pet dataset, ResNet18 demonstrated the best classification performance with metric values of 0.9986, 1, 1, and 1 for Accuracy, F1_Score, Precision, and Recall, respectively. In the case of the Food101 dataset, EfficientNetB0 achieved the highest classification performance, with values of 0.9320, 0.93, 0.94, and 0.93 for Accuracy, F1_Score, Precision, and Recall, respectively.