AISTATS 2023poster3 citations

Blessing of Class Diversity in Pre-training

Yulai Zhao, Jianshu Chen, Simon Du

Abstract

This paper presents a new statistical analysis aiming to explain the recent superior achievements of the pre-training techniques in natural language processing (NLP). We prove that when the classes of the pre-training task (e.g., different words in the masked language model task) are sufficiently diverse, in the sense that the least singular value of the last linear layer in pre-training (denoted as $\tilde{\nu}$) is large, then pre-training can significantly improve the sample efficiency of downstream tasks. Specially, we show the transfer learning excess risk enjoys an $O\left(\frac{1}{\tilde{\nu} \sqrt{n}}\right)$ rate, in contrast to the $O\left(\frac{1}{\sqrt{m}}\right)$ rate in the standard supervised learning. Here, $n$ is the number of pre-training data and $m$ is the number of data in the downstream task, and typically $n \gg m$. Our proof relies on a vector-form Rademacher complexity chain rule for disassembling composite function classes and a modified self-concordance condition. These techniques can be of independent interest.

BibTeX
@InProceedings{pmlr-v206-zhao23a,
  title = 	 {Blessing of Class Diversity in Pre-training},
  author =       {Zhao, Yulai and Chen, Jianshu and Du, Simon},
  booktitle = 	 {Proceedings of The 26th International Conference on Artificial Intelligence and Statistics},
  pages = 	 {283--305},
  year = 	 {2023},
  editor = 	 {Ruiz, Francisco and Dy, Jennifer and van de Meent, Jan-Willem},
  volume = 	 {206},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {25--27 Apr},
  publisher =    {PMLR},
  pdf = 	 {https://proceedings.mlr.press/v206/zhao23a/zhao23a.pdf},
  url = 	 {https://proceedings.mlr.press/v206/zhao23a.html},
  abstract = 	 {This paper presents a new statistical analysis aiming to explain the recent superior achievements of the pre-training techniques in natural language processing (NLP). We prove that when the classes of the pre-training task (e.g., different words in the masked language model task) are sufficiently diverse, in the sense that the least singular value of the last linear layer in pre-training (denoted as $\tilde{\nu}$) is large, then pre-training can significantly improve the sample efficiency of downstream tasks. Specially, we show the transfer learning excess risk enjoys an $O\left(\frac{1}{\tilde{\nu} \sqrt{n}}\right)$ rate, in contrast to the $O\left(\frac{1}{\sqrt{m}}\right)$ rate in the standard supervised learning. Here, $n$ is the number of pre-training data and $m$ is the number of data in the downstream task, and typically $n \gg m$. Our proof relies on a vector-form Rademacher complexity chain rule for disassembling composite function classes and a modified self-concordance condition. These techniques can be of independent interest.}
}