Large-scale pretrained language models are surprisingly good at recallingfactual knowledge presented in the training corpus. In this paper, we presentpreliminary studies on how factual knowledge is stored in pretrainedTransformers by introducing the concept of knowledge neurons. Specifically, weexamine the fill-in-the-blank cloze task for BERT. Given a relational fact, wepropose a knowledge attribution method to identify the neurons that express thefact. We find that the activation of such knowledge neurons is positivelycorrelated 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 onunderstanding the storage of knowledge within pretrained Transformers. The codeis available at https://github.com/Hunter-DDM/knowledge-neurons.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)