Journal article · 2026 · Open access

A Simple but Efficient Method to Achieve High Performance with Extreme Sparsity

Debin Liu, Xiang Bai, Ruonan Zhao, Xianjun Deng, Shibo He, Laurence Tianruo Yang

IEEE Transactions on Pattern Analysis and Machine Intelligence, PP, 1–15 · Published 12 November 2026 · DOI 10.1109/tpami.2026.3753244

Open full textCite this paper

In brief

“A Simple but Efficient Method to Achieve High Performance with Extreme Sparsity” is a 2026 journal article by Debin Liu and 5 co-authors, published in IEEE Transactions on Pattern Analysis and Machine Intelligence. OpenAlex counts 0 citations as of 18 November 2026. It is open access under a CC BY licence: the full text is free to read at IEEE Transactions on Pattern Analysis and Machine Intelligence.

Citations
0
Authors
6

Abstract

Pruning effectively removes redundant parameters and computations from neural networks, reduces model storage space, and saves time, memory, and energy during the training and inference phases. However, current pruning methods perform well at moderate and high sparsity, but the performance difference with dense network models at extreme sparsity is huge. In order to bridge the accuracy gap between extremely sparse and dense models, this paper proposes a simple but efficient pruning strategy-Tensor Decomposition Based Iterative Magnitude Pruning (TDIMP) method. This strategy uses tensor theory to optimize the model structure before IMP to reduce the search space of the sparse subnetwork during pruning, thus improving the hit rate of the optimal sparse subnetwork. Specifically, firstly, the weight matrix or weight tensor of the dense model is represented as a low-rank tensor decomposition format, and the feature extraction capability of the base module is enhanced by multi-step chaining operations to construct a low-rank compact network model. Secondly, IMP is performed on the low-rank model. The weight updates are sampled multiple times at the later stages of model training during each magnitude pruning process to smooth the oscillations of the learning trajectory and improve the performance of the sparse model. To evaluate the superiority of TDIMP, we first conducted extensive experiments using LSTM, VGGNet, ResNet, and Transformer architectures on UCF11, CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet datasets at sparsities ranging from 98% to 99.9%, and compared it with state-of-the-art methods. Experimental results show that TDIMP has higher accuracy, faster inference and less storage space than the pruning at initialization, dynamic sparse training, and iterative pruning methods in the extreme sparse case. In addition, TDIMP also has faster training speed than IMP methods.

Cite this paper

Formatted from the catalogue record. Check names, capitals and page numbers against the paper before you submit.

APA 7

Liu, D., Bai, X., Zhao, R., Deng, X., He, S., & Yang, L. T. (2026). A Simple but Efficient Method to Achieve High Performance with Extreme Sparsity. IEEE Transactions on Pattern Analysis and Machine Intelligence, PP, 1–15. https://doi.org/10.1109/tpami.2026.3753244

MLA 9

Liu, Debin, et al. “A Simple but Efficient Method to Achieve High Performance with Extreme Sparsity.” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PP, 2026, pp. 1–15. https://doi.org/10.1109/tpami.2026.3753244.

Chicago

Liu, Debin, Xiang Bai, Ruonan Zhao, Xianjun Deng, Shibo He, and Laurence Tianruo Yang. 2026. “A Simple but Efficient Method to Achieve High Performance with Extreme Sparsity.” IEEE Transactions on Pattern Analysis and Machine Intelligence PP: 1–15. https://doi.org/10.1109/tpami.2026.3753244.

Harvard

Liu, D. et al. (2026) ‘A Simple but Efficient Method to Achieve High Performance with Extreme Sparsity’, IEEE Transactions on Pattern Analysis and Machine Intelligence, PP, pp. 1–15. doi:10.1109/tpami.2026.3753244.

IEEE

D. Liu, X. Bai, R. Zhao, X. Deng, S. He, and L. T. Yang, “A Simple but Efficient Method to Achieve High Performance with Extreme Sparsity,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PP, pp. 1–15, 2026, doi: 10.1109/tpami.2026.3753244.

BibTeX

@article{liu2026simple,
  title     = {{A Simple but Efficient Method to Achieve High Performance with Extreme Sparsity}},
  author    = {Liu, Debin and Bai, Xiang and Zhao, Ruonan and Deng, Xianjun and He, Shibo and Yang, Laurence Tianruo},
  journal   = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year      = {2026},
  volume    = {PP},
  pages     = {1--15},
  doi       = {10.1109/tpami.2026.3753244},
  url       = {https://doi.org/10.1109/tpami.2026.3753244}
}

Download for a reference manager: BibTeX (.bib) · RIS for Zotero, Mendeley and EndNote (.ris)

Topic and keywords

Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications

  • neural network pruning
  • tensor decomposition
  • dynamic sparse training
  • inference efficiency

Work with this paper

Most-cited papers in Computer Science

All highly cited Computer Science papers

Questions about this paper

How do I cite this paper in APA style?

Liu, D., Bai, X., Zhao, R., Deng, X., He, S., & Yang, L. T. (2026). A Simple but Efficient Method to Achieve High Performance with Extreme Sparsity. IEEE Transactions on Pattern Analysis and Machine Intelligence, PP, 1–15. https://doi.org/10.1109/tpami.2026.3753244

How many times has this paper been cited?

0 times, according to OpenAlex on 18 November 2026. Google Scholar and Scopus count different sets of sources, so their numbers differ.

Is this paper free to read?

Yes. It is open access at IEEE Transactions on Pattern Analysis and Machine Intelligence: sign in and open the full text from this page.

Who wrote this paper?

Debin Liu, Xiang Bai, Ruonan Zhao, Xianjun Deng, Shibo He, Laurence Tianruo Yang.

What is the DOI of this paper?

10.1109/tpami.2026.3753244, permanently at https://doi.org/10.1109/tpami.2026.3753244. Sign in and open it from this page.

Bibliographic data from OpenAlex (CC0), last checked 18 November 2026. ResearchGuru does not host this paper; once you are signed in, the links open it at the publisher or the repository that does.