Gradient Localization Improves Lifelong Pretraining of Language Models
Preprint
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 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.Keywords
All Related Versions
This publication has 13 references indexed in Scilit:
- Continual pre-training mitigates forgetting in language and visionNeural Networks, 2024
- Meta-Learning Online Adaptation of Language Models2023
- Salient Span Masking for Temporal Understanding2023
- Time-Aware Language Models as Temporal Knowledge BasesTransactions of the Association for Computational Linguistics, 2022
- Fine-tuned Language Models are Continual Learners2022
- TimeLMs: Diachronic Language Models from Twitter2022
- Editing Factual Knowledge in Language Models2021
- Analyzing the Forgetting Problem in Pretrain-Finetuning of Open-domain Dialogue Response Models2021
- How Much Knowledge Can You Pack Into the Parameters of a Language Model?2020
- Recall and Learn: Fine-tuning Deep Pretrained Language Models with Less Forgetting2020