Evaluation of Few-Shot Methods from the Perspective of Class Scaling

Modern artificial intelligence models in computer vision applications can consist of billions of parameters, and their training may take several days. This entails the need for vast amounts of training data and significant energy consumption during training process. In the context of real-world business problems, there is a growing demand for models that are more flexible, faster, and have lower hardware requirements. In response to these needs, recent research has focused on developing universal methods that do not require costly retraining. Few-shot learning models have gained considerable interest due to their efficiency. Most of them emphasize the use of high-quality image feature extractors. However, an important yet often overlooked aspect of these methods is their scalability with respect to the increasing number of classes in classification tasks. Existing research rarely addresses this issue, instead focusing on achieving top results on the most popular benchmarks. This thesis aims to analyze and compare selected state-of-the-art few-shot learning methods in terms of their ability to scale with the number of classes in image classification problems. It contrasts large models with approaches that use feature extractors and classic clustering algorithms, highlighting the significant impact of the chosen extractor on final performance.