Keywords: natural language processing, adversarial training, mixed-strategy game, fine-tuning
TL;DR: We generalize fine-tuning adversarial training to a mixed-strategy game.
Abstract: Fine-tuning large-scale models from pre-trained checkpoints has been demonstrated effective for various natural language processing (NLP) tasks. Previous works reveal that leveraging adversarial training methods during the fine-tuning stage significantly enhances the generalization and robustness of the models. However, from the perspective of optimization, the previous adversarial training methods suffer from converging onto local optima due to the non-convexity of the objective. In this work, we reformulate the adversarial training in the view of mixed strategy in game theory and incorporate full strategy space to avoid trapping in local stationarity. Methodologically, we derive the Nash equilibrium of mixed-strategy for adversarial training using entropy mirror descent to establish a novel mixed-strategy adversarial training algorithm (MAT). Numerically, to verify the effectiveness of MAT, we conducted extensive benchmark experiments over the large-scale pre-trained models such as BERT and RoBERTa. The experimental results show that MAT outperforms the previous state-of-the-art on both GLUE and ANLI benchmarks in terms of generalization and robustness.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics
Submission Guidelines: Yes
Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning
Supplementary Material: zip