Better Representations via Adversarial Training in Pre-Training: A Theoretical Perspective
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2024
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866914655574687744 |
|---|---|
| author | Xing, Yue Lin, Xiaofeng Song, Qifan Xu, Yi Zeng, Belinda Cheng, Guang |
| author_facet | Xing, Yue Lin, Xiaofeng Song, Qifan Xu, Yi Zeng, Belinda Cheng, Guang |
| contents | Pre-training is known to generate universal representations for downstream tasks in large-scale deep learning such as large language models. Existing literature, e.g., \cite{kim2020adversarial}, empirically observe that the downstream tasks can inherit the adversarial robustness of the pre-trained model. We provide theoretical justifications for this robustness inheritance phenomenon. Our theoretical results reveal that feature purification plays an important role in connecting the adversarial robustness of the pre-trained model and the downstream tasks in two-layer neural networks. Specifically, we show that (i) with adversarial training, each hidden node tends to pick only one (or a few) feature; (ii) without adversarial training, the hidden nodes can be vulnerable to attacks. This observation is valid for both supervised pre-training and contrastive learning. With purified nodes, it turns out that clean training is enough to achieve adversarial robustness in downstream tasks. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2401_15248 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Better Representations via Adversarial Training in Pre-Training: A Theoretical Perspective Xing, Yue Lin, Xiaofeng Song, Qifan Xu, Yi Zeng, Belinda Cheng, Guang Machine Learning Pre-training is known to generate universal representations for downstream tasks in large-scale deep learning such as large language models. Existing literature, e.g., \cite{kim2020adversarial}, empirically observe that the downstream tasks can inherit the adversarial robustness of the pre-trained model. We provide theoretical justifications for this robustness inheritance phenomenon. Our theoretical results reveal that feature purification plays an important role in connecting the adversarial robustness of the pre-trained model and the downstream tasks in two-layer neural networks. Specifically, we show that (i) with adversarial training, each hidden node tends to pick only one (or a few) feature; (ii) without adversarial training, the hidden nodes can be vulnerable to attacks. This observation is valid for both supervised pre-training and contrastive learning. With purified nodes, it turns out that clean training is enough to achieve adversarial robustness in downstream tasks. |
| title | Better Representations via Adversarial Training in Pre-Training: A Theoretical Perspective |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2401.15248 |