New Insight of Variance reduce in Zero-Order Hard-Thresholding: Mitigating Gradient Error and Expansivity Contradictions
A new paper presents insights into variance reduction in zero-order hard-thresholding algorithms. The proposed method addresses limitations in existing algorithms by improving convergence rates and applicability. This research could enhance performance in machine learning tasks involving sparsity constraints.
- ▪The paper introduces a generalized variance reduced zero-order hard-thresholding algorithm.
- ▪This new approach mitigates conflicts between zero-order gradients and hard-thresholding operators.
- ▪The theoretical results indicate improved convergence rates compared to the existing SZOHT algorithm.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.18035 (cs) [Submitted on 18 May 2026] Title:New Insight of Variance reduce in Zero-Order Hard-Thresholding: Mitigating Gradient Error and Expansivity Contradictions Authors:Xinzhe Yuan (1), William de Vazelhes (2), Bin Gu (2 and 3), Huan Xiong (1 and 2) ((1) Harbin Institute of Technology, (2) Mohamed bin Zayed University of Artificial Intelligence, (3) Jilin University) View a PDF of the paper titled New Insight of Variance reduce in Zero-Order Hard-Thresholding: Mitigating Gradient Error and Expansivity Contradictions, by Xinzhe Yuan (1) and 5 other authors View PDF HTML (experimental) Abstract:Hard-thresholding is an important type of algorithm in machine learning that is used to solve $\ell_0$ constrained optimization problems.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
