On Transfer in Classification: How Well do Subsets of Classes Generalize?
Résumé
In classification settings, models trained on certain classes may apply to unseen ones, indicating learning beyond the original task. This concept is frequently utilized in transfer learning, where pretrained models can handle new classes. However, the theoretical underpinnings of this phenomenon remain largely unexplored. This study aims to establish a theoretical framework for class transferability. We introduce a partial order relation for class sets, enabling the characterization of class sets that can generalize to others. In practical terms, we investigate our framework's capacity to predict which class sets yield optimal performance when tested on new classes. Additionally, we delve into fewshot learning, where transfer is essential. Our work contributes to better understanding of transfer mechanics and model generalization.