A Closer Look at How Fine-tuning Changes BERT

TL;DR

This paper investigates how fine-tuning BERT alters its embedding space, revealing that it mainly adjusts representations to specific tasks while maintaining their original structure, and challenges the assumption that fine-tuning always improves performance.

Contribution

It provides a detailed analysis of the effects of fine-tuning on BERT's embedding space using probing techniques and experiments across multiple NLP tasks.

Findings

Fine-tuning increases distances between differently labeled examples.

Fine-tuning does not significantly distort the original embedding structure.

There exists an exception where fine-tuning does not improve performance.

Abstract

Given the prevalence of pre-trained contextualized representations in today's NLP, there have been many efforts to understand what information they contain, and why they seem to be universally successful. The most common approach to use these representations involves fine-tuning them for an end task. Yet, how fine-tuning changes the underlying embedding 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. We hypothesize that fine-tuning affects classification performance by increasing the distances between examples associated with different labels. We confirm this hypothesis with carefully designed experiments on five different NLP tasks. Via these experiments, we also discover an exception to the prevailing wisdom that "fine-tuning always improves performance". Finally, by comparing the representations before and after fine-tuning, we discover that fine-tuning does not introduce arbitrary changes to representations; instead, it adjusts the representations to downstream tasks while largely preserving the original spatial structure of the data points.