Xiao Liang 的照片

博士生 · 爱荷华州立大学

梁笑

Xiao Liang

一个鞍点接一个鞍点，绘制非凸优化的地形图。

电气与计算机工程系
爱荷华州立大学 · Ames, Iowa

关于我

你好，我是梁笑！目前在爱荷华州立大学电气与计算机工程系读博士，导师是 Shuang Li 教授。

一直吸引我的，是一个听上去简单得有点过分的问题：有些非凸问题在纸面上看起来毫无希望，为什么最朴素的梯度下降偏偏能把它们解得很好？为了找到答案，我去刻画优化问题的“地形”：在结构化的低秩问题里寻找山谷、山脊和鞍点，比如只凭幅值测量同时恢复信号和模糊核，或是借助傅里叶切片研究张量感知。把几何结构弄清楚之后，我也喜欢把这些理解变成更快的算法。最近在研究 Muon 式的正交化更新，看它能把张量回归加速多少。

我的求学路线拐过几个弯：本科在北京的北方工业大学读数据科学（2018–2022），之后在圣路易斯华盛顿大学读计算机科学（2022–2024），读博前又在华为实习了一段时间，然后来到 Ames。爱荷华是出了名的一马平川，所以现在真正有起伏的地形，都在我的损失函数里。

试一试 在图上任意位置点一下，放下一个小球，看它沿着梯度一路滚下去。中间有一个鞍点，但没有伪局部极小值，所以每个小球最后都会落到两个全局最优解 ±x★ 中的一个。这种正负号上的不确定性，正是相位恢复必须面对的问题。

研究

  1. r₁=2 r₂=2 r₃=1 r₄=3 strict saddle Preprint

    Strict-Saddle Landscapes and Multi-Rank Geometry in Low-Tubal-Rank Tensor Sensing

    Eugene Agyei-Kodie*, Longxiu Huang*, Shuang Li†, Xiao Liang†

    arXiv preprint, 2026 (*, † equal contribution)

    Under a tubal restricted isometry condition, the balanced factorized objective has no spurious local minima and a quantitative strict-saddle structure. Its local geometry is governed by the ranks of the Fourier slices, not by the tubal rank alone.

    arXiv PDF
    BibTeX 
    @article{agyeikodie2026strict,
  title   = {Strict-Saddle Landscapes and Multi-Rank Geometry in Low-Tubal-Rank Tensor Sensing},
  author  = {Agyei-Kodie, Eugene and Huang, Longxiu and Li, Shuang and Liang, Xiao},
  journal = {arXiv preprint arXiv:2609.37865},
  year    = {2026}
}
  2. U ← NS(momentum) Preprint

    A Muon-Accelerated Algorithm for Low Separation Rank Tensor Generalized Linear Models

    Xiao Liang, Shuang Li

    arXiv preprint, 2026

    LSRTR-M replaces the QR-based projection steps of LSRTR with Muon (momentum orthogonalized by Newton–Schulz) updates. On linear, logistic, and Poisson tensor GLMs it converges faster in both iterations and wall-clock time, and it stays competitive on Vessel MNIST 3D classification.

    arXiv PDF Code
    BibTeX 
    @article{liang2026muon,
  title   = {A Muon-Accelerated Algorithm for Low Separation Rank Tensor Generalized Linear Models},
  author  = {Liang, Xiao and Li, Shuang},
  journal = {arXiv preprint arXiv:2604.04726},
  year    = {2026}
}
  3. −x★ +x★ ICASSP 2026

    Landscape Analysis of Simultaneous Blind Deconvolution and Phase Retrieval

    Xiao Liang, Zhen Qin, Zhihui Zhu, Shuang Li

    IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Barcelona, 2026

    Blind deconvolution and phase retrieval, solved at the same time, become a structured low-rank tensor recovery problem. For a tractable surrogate, the paper characterizes the global landscape of the population risk on the sphere and shows that Riemannian gradient descent converges linearly.

    Paper DOI
    BibTeX 
    @inproceedings{liang2026landscape,
  title     = {Landscape Analysis of Simultaneous Blind Deconvolution and Phase Retrieval},
  author    = {Liang, Xiao and Qin, Zhen and Zhu, Zhihui and Li, Shuang},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages     = {511--515},
  year      = {2026},
  doi       = {10.1109/ICASSP55912.2026.11461778}
}
  4. x★ noise robust Preprint

    Landscape Analysis of Simultaneous Blind Deconvolution and Phase Retrieval via Structured Low-Rank Tensor Recovery

    Xiao Liang*, Zhen Qin*, Zhihui Zhu, Shuang Li

    arXiv preprint, 2025 (* equal contribution)

    The full story behind the ICASSP paper. Beyond the global landscape of the population surrogate, it establishes local geometry, convergence guarantees, and robustness to measurement noise for the full tensor sensing problem, with experiments that point the way back to the original problem.

    arXiv PDF
    BibTeX 
    @article{liang2025landscape,
  title   = {Landscape Analysis of Simultaneous Blind Deconvolution and Phase Retrieval via Structured Low-Rank Tensor Recovery},
  author  = {Liang, Xiao and Qin, Zhen and Zhu, Zhihui and Li, Shuang},
  journal = {arXiv preprint arXiv:2509.10834},
  year    = {2025}
}

其他论文

教育经历