This is the repository for the COLM 24 paper: Does RoBERTa Perform Better than BERT in Continual Learning: An Attention Sink Perspective.
This project studies the Continual learning (CL) problem with pre-trained language models.
- Pe-trained models may allocate high attention scores to some 'sink' tokens, such as [SEP] tokens, which are ubiquitous across various tasks. Such attention sinks may lead to models' over-smoothing in single-task learning and interference in sequential tasks’ learning, downgrading models' CL capacities.
- We propose a pre-scaling mechanism that encourages attention diversity across all tokens. Specifically, it first scales the task's attention to the non-sink tokens in a probing stage, and then fine-tunes the model with scaling.
- Experiments show that pre-scaling yields substantial improvements in CL without experience replay, or progressively storing parameters from previous tasks.
numpy == 1.16.2
torch == 1.9.1
transformers == 3.0.0
Each task data are stored in the ./data directory. For tasks in GLUE benchmark, please download the data in this link to the corresponding sub directories. Other data can be downloaded here.
To preprocess the data, please run python preprocess.py in each sub directory. For Yahoo and DBPedia which need to be split to sub-class tasks, please run python preprocess_split.py.
The scripts for running BERT and RoBERTa with pre-scaling is in ./bert and ./roberta directory, respectively.
For BERT models, we show sample commands for prescaling in ./bert/script_prescale.sh and for probing-and-then-fine-tuning in ./bert/script_prescale.sh. And for RoBERTa models we show similar .sh files.
@inproceedings{
bai2024does,
title={Does Ro{BERT}a Perform Better than {BERT} in Continual Learning: An Attention Sink Perspective},
author={Xueying Bai and Yifan Sun and Niranjan Balasubramanian},
booktitle={First Conference on Language Modeling},
year={2024},
url={https://openreview.net/forum?id=VHhwhmtx3b}
}