Few and Fewer: Learning Better from Few Examples Using Fewer Base Classes

Few-shot learning considers problems where training data is severely limited. It represents a challenge for deep learning, which typically requires large datasets of training examples (Wang et al., 2020b). The standard technique to leverage deep learning within few-shot tasks is to adopt some form of transfer learning, using a large distinct "base dataset" to train a model, which then serves as a feature extractor to integrate additional knowledge within the "target dataset", on which the task has to be solved. One of the most straightforward transfer strategies is thus to embed the target data into an appropriate feature space, and then to learn a simple classifier with minimal parameters in order to avoid overfitting to the few labeled examples (Wang et al., 2019). However, the effectiveness of the transfer depends on the similarity of the base and target domains, and recent research in transfer learning (Oh et al., 2022) suggests that it may even have a deleterious effect if the domain gap is too large (Guo et al., 2020). This paper therefore considers the question: can one reduce the domain gap by fine-tuning on the base classes that are the most similar to the target distribution?. This approach aims to minimize the domain gap by concentrating the model's learning on a narrower, more relevant subset of the base classes closely aligned with the target distribution. This questions the existence of universal feature extractors that would lead to systematically high performance on any few-shot task, a common trend in the field (Kirillov et al., 2023).