English

Pruning Redundant Mappings in Transformer Models via Spectral-Normalized Identity Prior

Computation and Language 2020-10-06 v1

Abstract

Traditional (unstructured) pruning methods for a Transformer model focus on regularizing the individual weights by penalizing them toward zero. In this work, we explore spectral-normalized identity priors (SNIP), a structured pruning approach that penalizes an entire residual module in a Transformer model toward an identity mapping. Our method identifies and discards unimportant non-linear mappings in the residual connections by applying a thresholding operator on the function norm. It is applicable to any structured module, including a single attention head, an entire attention block, or a feed-forward subnetwork. Furthermore, we introduce spectral normalization to stabilize the distribution of the post-activation values of the Transformer layers, further improving the pruning effectiveness of the proposed methodology. We conduct experiments with BERT on 5 GLUE benchmark tasks to demonstrate that SNIP achieves effective pruning results while maintaining comparable performance. Specifically, we improve the performance over the state-of-the-art by 0.5 to 1.0% on average at 50% compression ratio.

Keywords

Cite

@article{arxiv.2010.01791,
  title  = {Pruning Redundant Mappings in Transformer Models via Spectral-Normalized Identity Prior},
  author = {Zi Lin and Jeremiah Zhe Liu and Zi Yang and Nan Hua and Dan Roth},
  journal= {arXiv preprint arXiv:2010.01791},
  year   = {2020}
}

Comments

Findings of EMNLP 2020

R2 v1 2026-06-23T19:01:49.210Z