A Closer Look at How Fine-tuning Changes BERT

Given the prevalence of pre-trained contextualized representations in today's NLP, there have been several efforts to understand what information such representations contain. A common strategy to use such representations is to fine-tune them for an end task. However, how fine-tuning for a task changes the underlying space is less studied. In this work, we study the English BERT family and use two probing techniques to analyze how fine-tuning changes the space. Our experiments reveal that fine-tuning improves performance because it pushes points associated with a label away from other labels. By comparing the representations before and after fine-tuning, we also discover that fine-tuning does not change the representations arbitrarily; instead, it adjusts the representations to downstream tasks while preserving the original structure. Finally, using carefully constructed experiments, we show that fine-tuning can encode training sets in a representation, suggesting an overfitting problem of a new kind.