Gradient Localization Improves Lifelong Pretraining of Language Models
This addresses the challenge of maintaining and updating knowledge in large language models for applications requiring up-to-date information, but it is incremental as it builds on existing continual learning methods.
The paper tackled the problem of catastrophic forgetting and failed uptake of new information in continual pretraining of language models by showing that knowledge is localized to specific parameters, and demonstrated that targeting updates to relevant layers improves performance on temporal drift data.
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 contributes to both: the failed uptake of new information, and catastrophic forgetting of previously learned information. We observe that sequences containing references to updated and newly mentioned entities exhibit larger gradient norms in a subset of layers. We demonstrate that targeting parameter updates to these relevant layers can improve the performance of continually pretraining on language containing temporal drift.