Large-scale pretrained language models are surprisingly good at recalling factual knowledge presented in the training corpus. In this paper, we present preliminary studies on how factual knowledge is stored in pretrained Transformers by introducing the concept of knowledge neurons. Specifically, we examine the fill-in-the-blank cloze task for BERT. Given a relational fact, we propose a knowledge attribution method to identify the neurons that express the fact. We find that the activation of such knowledge neurons is positively correlated to the expression of their corresponding facts. In our case studies, we attempt to leverage knowledge neurons to edit (such as update, and erase) specific factual knowledge without fine-tuning. Our results shed light on understanding the storage of knowledge within pretrained Transformers.

The recently introduced pre-trained language model BERT advances the state-of-the-art on many NLP tasks through the fine-tuning approach, but few studies investigate how the fine-tuning process improves the model performance on downstream tasks. In this paper, we inspect the learning dynamics of BERT fine-tuning with two indicators. We use JS divergence to detect the change of the attention mode and use SVCCA distance to examine the change to the feature extraction mode during BERT fine-tuning. We conclude that BERT fine-tuning mainly changes the attention mode of the last layers and modifies the feature extraction mode of the intermediate and last layers. Moreover, we analyze the consistency of BERT fine-tuning between different random seeds and different datasets. In summary, we provide a distinctive understanding of the learning dynamics of BERT fine-tuning, which sheds some light on improving the fine-tuning results.

Language model pre-training, such as BERT, has achieved remarkable results in many NLP tasks. However, it is unclear why the pre-training-then-fine-tuning paradigm can improve performance and generalization capability across different tasks. In this paper, we propose to visualize loss landscapes and optimization trajectories of fine-tuning BERT on specific datasets. First, we find that pre-training reaches a good initial point across downstream tasks, which leads to wider optima and easier optimization compared with training from scratch. We also demonstrate that the fine-tuning procedure is robust to overfitting, even though BERT is highly over-parameterized for downstream tasks. Second, the visualization results indicate that fine-tuning BERT tends to generalize better because of the flat and wide optima, and the consistency between the training loss surface and the generalization error surface. Third, the lower layers of BERT are more invariant during fine-tuning, which suggests that the layers that are close to input learn more transferable representations of language.