Best Papers
 Sparsification and Sparsity-Preserving PSD Correction for Cross-Layer Aware Mixed-Precision Quantization 


Vol. 51,  No. 4, pp. 884-894, Apr.  2026
10.7840/kics.2026.51.4.884


PDF Full-Text
  Abstract

Mixed-precision quantization (MPQ) is a key technique for deep learning model compression. While recent methods that consider cross-layer dependencies have achieved high performance, their approaches to correcting non-Positive Semi-Definite (PSD) sensitivity matrices introduce fundamental limitations, such as distorting the optimal solution and destroying matrix sparsity. In this paper, we propose a novel MPQ framework. Our proposed method applies sensitivity matrix sparsification by eliminating unnecessary long-range layer interactions. Also, we developed a new PSD correction method that preserves the induced sparse structure and achieves stable optimization without distorting the optimal solution. Experimental results on the ImageNet-1K dataset with a ResNet-34 model demonstrate that our framework achieves higher Top-1 accuracy compared to existing cross-layer dependency-based methods.

  Statistics
Cumulative Counts from November, 2022
Multiple requests among the same browser session are counted as one view. If you mouse over a chart, the values of data points will be shown.


  Related Articles
  Cite this article

[IEEE Style]

J. Lee, M. Jung, K. Son, "Sparsification and Sparsity-Preserving PSD Correction for Cross-Layer Aware Mixed-Precision Quantization," The Journal of Korean Institute of Communications and Information Sciences, vol. 51, no. 4, pp. 884-894, 2026. DOI: 10.7840/kics.2026.51.4.884.

[ACM Style]

JHye-seong Lee, Min-kyu Jung, and Kyungrak Son. 2026. Sparsification and Sparsity-Preserving PSD Correction for Cross-Layer Aware Mixed-Precision Quantization. The Journal of Korean Institute of Communications and Information Sciences, 51, 4, (2026), 884-894. DOI: 10.7840/kics.2026.51.4.884.

[KICS Style]

JHye-seong Lee, Min-kyu Jung, Kyungrak Son, "Sparsification and Sparsity-Preserving PSD Correction for Cross-Layer Aware Mixed-Precision Quantization," The Journal of Korean Institute of Communications and Information Sciences, vol. 51, no. 4, pp. 884-894, 4. 2026. (https://doi.org/10.7840/kics.2026.51.4.884)
Vol. 51, No. 4 Index

Search
(IN TITLE, AUTHOR, ABSTRACT,KEYWORDS)

Advanced Search ▼

POPULAR KEYWORDS
(TOP 10 KEYWORDS)

Recent Publications
(LAST 3 YEARS)






pISSN : 1226 - 4717
eISSN : 2287 - 3880