Source author record

Pengyue Hou

Pengyue Hou appears in the imported research catalog. Authorship, coauthor and topic links are available while profile ownership is still unclaimed.

ResearcherUnclaimed source record

Catalog footprint

What is connected

2works
6topics
4close collaborators

Actions

Connect this record

Log in to claim

Research graph

See the researcher in context

Open full explorer

Inspect adjacent papers, topics, institutions and collaborators without losing the researcher page.

Building this map preview

BZPEER is loading the nearby papers, people, topics and institutions for this page.

Published work

2 published item(s)

preprint2026arXiv

Global Dynamics and Stabilization of Zero-Mode Singularities in Multi-Scale Reaction-Diffusion Systems via Negative Coupling

This paper establishes a rigorous mathematical framework for the Multi-Scale Negative Coupled System (MNCS), a dynamical model describing hierarchical state spaces with directed, sign-structured interactions. We address the stabilization of reaction-diffusion systems on bounded domains $Ω\subset \mathbb{R}^d$ ($d \le 3$) subject to homogeneous Neumann boundary conditions. A critical feature of this setting is the "zero-mode singularity," where the Laplacian operator possesses a trivial zero eigenvalue ($λ_0=0$), providing no linear dissipation for the spatial mean. We rigorously prove the global well-posedness of the system and the existence of a compact global attractor $\mathcal{A}$ in the phase space $\mathbb{H}=(L^2(Ω))^N$. Utilizing the Moser-Alikakos iteration technique, we establish uniform $L^\infty(Ω)$ bounds, overcoming the lack of Sobolev embedding from $H^1$ into $L^\infty$ in three dimensions. These bounds enable the derivation of explicit upper estimates for the fractal dimension of the attractor via the Kaplan-Yorke trace formula. We show that the dimension scales as $d_F(\mathcal{A}) \sim \max\{0, \mathcal{K}_{\mathcal{A}}-γ\}^{d/2}$, confirming that the negative coupling strength $γ$ acts as a global regularizer that compresses the phase space. The theoretical results are validated using a stiff-stable Second-Order Exponential Time Differencing (ETD2) scheme with Discrete Cosine Transform (DCT) to strictly enforce no-flux boundary conditions.

preprint2022arXiv

Adversarial Fine-tune with Dynamically Regulated Adversary

Adversarial training is an effective method to boost model robustness to malicious, adversarial attacks. However, such improvement in model robustness often leads to a significant sacrifice of standard performance on clean images. In many real-world applications such as health diagnosis and autonomous surgical robotics, the standard performance is more valued over model robustness against such extremely malicious attacks. This leads to the question: To what extent we can boost model robustness without sacrificing standard performance? This work tackles this problem and proposes a simple yet effective transfer learning-based adversarial training strategy that disentangles the negative effects of adversarial samples on model's standard performance. In addition, we introduce a training-friendly adversarial attack algorithm, which facilitates the boost of adversarial robustness without introducing significant training complexity. Extensive experimentation indicates that the proposed method outperforms previous adversarial training algorithms towards the target: to improve model robustness while preserving model's standard performance on clean data.