Gradient Localization Improves Lifelong Pretraining of Language Models
Abstract: Large Language Models (LLMs) trained on web-scale text corpora have been shown to capture world knowledge in their parameters. However, the mechanism by which language models store different types of knowledge is poorly understood. In this work, we examine two types of knowledge relating to temporally sensitive entities and demonstrate that each type is localized to different sets of parameters within the LLMs. We hypothesize that the lack of consideration of the locality of knowledge in existing continual learning methods is responsible for failed uptake of new information and catastrophic forgetting of previously learned information. We demonstrate that targeted training to these relevant layers can improve the performance of continually learned language under temporal drift.
Paper Type: Short
Research Area: Machine Learning for NLP
Research Area Keywords: gradient norm, continual learning, continual pretraining, temporal
Contribution Types: Model analysis & interpretability
Languages Studied: English
Submission Number: 5870
Loading