Better Fine-Tuning by Reducing Representational Collapse
arXiv:2008.03156
Abstract
Although widely adopted, existing approaches for fine-tuning pre-trained language models have been shown to be unstable across hyper-parameter settings, motivating recent work on trust region methods. In this paper, we present a simplified and efficient method rooted in trust region theory that replaces previously used adversarial objectives with parametric noise (sampling from either a normal or uniform distribution), thereby discouraging representation change during fine-tuning when possible without hurting performance. We also introduce a new analysis to motivate the use of trust region methods more generally, by studying representational collapse; the degradation of generalizable representations from pre-trained models as they are fine-tuned for a specific end task. Extensive experiments show that our fine-tuning method matches or exceeds the performance of previous trust region methods on a range of understanding and generation tasks (including DailyMail/CNN, Gigaword, Reddit TIFU, and the GLUE benchmark), while also being much faster. We also show that it is less prone to representation collapse; the pre-trained models maintain more generalizable representations every time they are fine-tuned.
References in corpus (9)
- Trust Region Policy Optimization
- Teaching Machines to Read and Comprehend
- PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization
- FreeLB: Enhanced Adversarial Training for Natural Language Understanding
- On Mutual Information Maximization for Representation Learning
- Fine-Tuning Pretrained Language Models: Weight Initializations, Data Orders, and Early Stopping
- Pre-training via Paraphrasing
- ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training
- InfoXLM: An Information-Theoretic Framework for Cross-Lingual Language Model Pre-Training