IsoBN: Fine-Tuning BERT with Isotropic Batch Normalization
This work addresses a specific bottleneck in fine-tuning language models for researchers and practitioners in NLP, offering a simple method to enhance performance, though it is incremental as it builds on existing fine-tuning practices.
The paper tackled the problem of anisotropic embeddings in pre-trained language models like BERT and RoBERTa, which can hinder performance in natural language understanding tasks, and proposed IsoBN, a regularization method that improved the average performance by about 1.0 absolute increment on seven NLU tasks.
Fine-tuning pre-trained language models (PTLMs), such as BERT and its better variant RoBERTa, has been a common practice for advancing performance in natural language understanding (NLU) tasks. Recent advance in representation learning shows that isotropic (i.e., unit-variance and uncorrelated) embeddings can significantly improve performance on downstream tasks with faster convergence and better generalization. The isotropy of the pre-trained embeddings in PTLMs, however, is relatively under-explored. In this paper, we analyze the isotropy of the pre-trained [CLS] embeddings of PTLMs with straightforward visualization, and point out two major issues: high variance in their standard deviation, and high correlation between different dimensions. We also propose a new network regularization method, isotropic batch normalization (IsoBN) to address the issues, towards learning more isotropic representations in fine-tuning by dynamically penalizing dominating principal components. This simple yet effective fine-tuning method yields about 1.0 absolute increment on the average of seven NLU tasks.