Improving language models fine-tuning with representation consistency targets

Anastasia, Razdaibiedina,; Madan, Vivek; Karnin, Zohar; Khetan, Ashish; Kapoor, Vishaal

doi:10.48550/arxiv.2205.11603

Cited by 1 publication

(1 citation statement)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Additionally, for many applications such storage would not be feasible due to privacy settings, when access to past data is not available. Regularization-based approaches are more memoryefficient than replay-based approaches, but suffer from catastrophic forgetting and are often not suitable for long task sequences (Kirkpatrick et al, 2017;Razdaibiedina et al, 2022). In contrast to regularization-based and replay-based approaches, architectural CL approaches are more efficient in resolving catastrophic forgetting and, hence, are suitable for long sequences of tasks.…”

Section: Forward Transfer Experimentsmentioning

confidence: 99%

Progressive Prompts: Continual Learning for Language Models

Anastasia¹,

Mao²,

Hou³

et al. 2023

Preprint

View full text Add to dashboard Cite

We introduce Progressive Prompts -a simple and efficient approach for continual learning in language models. Our method allows forward transfer and resists catastrophic forgetting, without relying on data replay or a large number of task-specific parameters. Progressive Prompts learns a new soft prompt for each task and sequentially concatenates it with the previously learned prompts, while keeping the base model frozen. Experiments on standard continual learning benchmarks show that our approach outperforms state-of-the-art methods, with an improvement >20% in average test accuracy over the previous best-preforming method on T5 model. We also explore a more challenging continual learning setup with longer sequences of tasks and show that Progressive Prompts significantly outperforms prior methods.

show abstract

Section: Forward Transfer Experimentsmentioning

confidence: 99%