Task-Agnostic Structured Pruning of Speech Representation Models

被引:1
|
作者
Wang, Haoyu [1 ]
Wang, Siyuan [1 ]
Zhang, Wei-Qiang [1 ]
Suo, Hongbin [2 ]
Wan, Yulong [2 ]
机构
[1] Tsinghua Univ, Dept Elect Engn, Beijing 100084, Peoples R China
[2] OPPO, Data & AI Engn Syst, Beijing 100026, Peoples R China
来源
基金
中国国家自然科学基金;
关键词
Model pruning; knowledge distillation; model compression; representation learning;
D O I
10.21437/Interspeech.2023-1442
中图分类号
O42 [声学];
学科分类号
070206 ; 082403 ;
摘要
Self-supervised pre-trained models such as Wav2vec2, Hubert, and WavLM have been shown to significantly improve many speech tasks. However, their large memory and strong computational requirements hinder their industrial applicability. Structured pruning is a hardware-friendly model compression technique but usually results in a larger loss of accuracy. In this paper, we propose a fine-grained attention head pruning method to compensate for the performance degradation. In addition, we also introduce the straight through estimator into the L0 regularization to further accelerate the pruned model. Experiments on the SUPERB benchmark show that our model can achieve comparable performance to the dense model in multiple tasks and outperforms the Wav2vec 2.0 base model on average, with 72% fewer parameters and 2 times faster inference speed.
引用
收藏
页码:231 / 235
页数:5
相关论文
共 50 条
  • [41] Task-Agnostic Privacy-Preserving Representation Learning for Federated Learning against Attribute Inference Attacks
    Arevalo, Caridad Arroyo
    Noorbakhsh, Sayedeh Leila
    Dong, Yun
    Hong, Yuan
    Wang, Binghui
    THIRTY-EIGHTH AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE, VOL 38 NO 10, 2024, : 10909 - 10917
  • [42] Investigating the Effectiveness of Task-Agnostic Prefix Prompt for Instruction Following
    Ye, Seonghyeon
    Hwang, Hyeonbin
    Yang, Sohee
    Yun, Hyeongu
    Kim, Yireun
    Seo, Minjoon
    THIRTY-EIGHTH AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE, VOL 38 NO 17, 2024, : 19386 - 19394
  • [43] A Self-supervised Task-agnostic Embedding for EEG Signals
    Partovi, Andi
    Burkitt, Anthony N.
    Grayden, David
    2023 11TH INTERNATIONAL IEEE/EMBS CONFERENCE ON NEURAL ENGINEERING, NER, 2023,
  • [44] Downstream Task-agnostic Transferable Attacks on Language-Image Pre-training Models
    Lv, Yiqiang
    Chen, Jingjing
    Wei, Zhipeng
    Chen, Kai
    Wu, Zuxuan
    Jiang, Yu-Gang
    2023 IEEE INTERNATIONAL CONFERENCE ON MULTIMEDIA AND EXPO, ICME, 2023, : 2831 - 2836
  • [45] Task-Agnostic Vision Transformer for Distributed Learning of Image Processing
    Kim, Boah
    Kim, Jeongsol
    Ye, Jong Chul
    IEEE TRANSACTIONS ON IMAGE PROCESSING, 2023, 32 : 203 - 218
  • [46] Extract then Distill: Efficient and Effective Task-Agnostic BERT Distillation
    Chen, Cheng
    Yin, Yichun
    Shang, Lifeng
    Wang, Zhi
    Jiang, Xin
    Chen, Xiao
    Liu, Qun
    ARTIFICIAL NEURAL NETWORKS AND MACHINE LEARNING - ICANN 2021, PT III, 2021, 12893 : 570 - 581
  • [47] Task-Agnostic Adaptation for Safe Human-Robot Handover
    Liu, Ruixuan
    Chen, Rui
    Liu, Changliu
    IFAC PAPERSONLINE, 2022, 55 (41): : 175 - 180
  • [48] Interesting Object, Curious Agent: Learning Task-Agnostic Exploration
    Parisi, Simone
    Dean, Victoria
    Pathak, Deepak
    Gupta, Abhinav
    ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 34 (NEURIPS 2021), 2021, 34
  • [49] CodePrompt: Task-Agnostic Prefix Tuning for Program and Language Generation
    Choi, YunSeok
    Lee, Jee-Hyong
    FINDINGS OF THE ASSOCIATION FOR COMPUTATIONAL LINGUISTICS, ACL 2023, 2023, : 5282 - 5297
  • [50] Continual deep reinforcement learning with task-agnostic policy distillation
    Hafez, Muhammad Burhan
    Erekmen, Kerim
    SCIENTIFIC REPORTS, 2024, 14 (01):