About

My name is Qizhe Yang. I obtained a PhD degree from the School of Software, Shanghai Jiao Tong University in 2022. My supervisor is Professor Yuxi Fu. I am now a lecturer at the College of Information, Mechanical and Electrical Engineering, Shanghai Normal University (SHNU). I am interested in understanding the behavior of complex systems from both theoretical and applied perspectives: on one hand, studying reachability, complexity, and decidability problems in infinite-state systems; on the other hand, exploring the reliability, interpretability, and generalization capability of computer vision and AI systems in complex environments.

Education

  • 2016.9 – 2022.6   PhD, Shanghai Jiao Tong University, supervised by Prof. Yuxi Fu.
  • 2012.9 – 2016.6   BSc, Shanghai Jiao Tong University.

Research

Reachability in Infinite-State Systems

Infinite-state systems are important objects of study in formal verification and theoretical computer science, widely used to model computational systems with counters, resource constraints, communication behavior, or concurrent structure. Among these, Vector Addition Systems with States (VASS) and their variants are classical models for studying reachability in infinite-state systems. The VASS reachability problem has deep theoretical significance and is closely related to program verification, concurrent system analysis, and Petri net theory.

I am currently focused on reachability problems in low-dimensional VASS and their extensions, with particular attention to the fine-grained relationship between dimension, zero-test capability, and complexity.

1. 3-VASS Reachability  (PSPACE lower bound · EXPSPACE upper bound · draft · joint work with BASICS members)

The 3-VASS reachability problem lies at the heart of complexity research on low-dimensional VASS. In 2023 I established a Tower upper bound for this problem, which was subsequently improved to 2-EXPSPACE in 2025. We currently have new progress towards an EXPSPACE upper bound (draft), but the exact complexity remains open.

2. d-VASS Reachability  (F(d−3)/2 lower bound · Fd upper bound · ICALP 2024)

For VASS of general fixed dimension, the complexity of the reachability problem grows rapidly with dimension. Known results show a significant gap between lower and upper bounds: F(d−3)/2 and Fd respectively. My interest lies in understanding the essential impact of dimension on reachability complexity: which behaviors inherently require high dimension? Can more refined structural decomposition and self-reduction techniques further improve the upper bounds?

3. 3-VASS0 Reachability  (decidability resolved · interested in the separation from 3-VASS)

3-VASS0 extends 3-VASS with a zero-test capability. Zero-testing significantly enhances expressive power, bringing the model closer to stronger computational systems. Although decidability is settled, the precise complexity-theoretic separation from ordinary 3-VASS remains worth deeper investigation. I am particularly interested in how a single zero-testable counter changes the model's capabilities.

Computer Vision

In computer vision, I focus on image restoration tasks — particularly the modeling and recovery of image degradation under adverse weather conditions. Image restoration is not only a fundamental low-level vision task but also directly affects the performance of downstream detection, recognition, segmentation, and multimodal perception systems.

Previous work proposed a unified modeling framework with physical and statistical consistency and designed the Multi-Component Decomposition Network (MCD-Net) for explicit decoupling and joint restoration of multiple degradation components. The goal is to provide solutions with stronger physical interpretability and generalization ability, and to supply high-quality pre-processing support for multi-task visual perception pipelines.

1. Rain Removal in Complex Weather Scenes

Traditional rain removal typically addresses daytime scenarios with relatively uniform degradation. In more complex environments — nighttime, strong reflections, low illumination, mixed rain and haze, dynamic light sources — image degradation exhibits far richer physical and statistical characteristics. For example, nighttime rainy images may simultaneously contain raindrops, halos, noise, overexposure, and low-light detail loss. I aim to study deraining models closer to real-world conditions, with particular attention to nighttime deraining, multi-degradation coupled modeling, and robustness under complex weather.

2. Pre-Generalization in Model Training

Many image restoration models perform well on specific datasets but degrade significantly when facing unknown degradation types or real-world distribution shifts. The core questions are: how to leverage complementary information from synthetic data, real data, and physical priors? How to train models that maintain stable restoration capability in unseen complex scenarios? This direction connects to domain generalization, foundation visual models, physical prior modeling, and self-supervised learning.

Trustworthy AI

Trustworthy AI is a direction I am actively exploring. As large models and multimodal models enter research, education, and industry, optimizing for performance alone is no longer sufficient — we need to understand why models work, when they fail, and how to use AI safely in high-reliability settings.

1. AI Interpretability

AI interpretability concerns understanding a model's decision basis, internal representations, and failure modes. In vision tasks this may mean identifying which image regions or texture patterns the model relies on; in large language models it may concern how reasoning steps are organized or how errors arise. I hope to study explainable AI from both theoretical and applied perspectives: exploring the relationship between internal representations, attention structures, and output behavior, and designing interpretable, diagnosable, intervenable AI systems.

2. A New Paradigm for AI-Assisted Research

The development of large models is changing how scientific research is organized. AI can participate not only in literature review, code generation, and experimental assistance, but also in deeper research processes: proposing conjectures, searching for counterexamples, assisting with proofs, formal verification, and experiment design. I am particularly interested in AI-assisted theoretical computer science — for example, whether AI can help discover new structural decompositions in VASS and complexity theory, or combine with formal tools to produce more reliable mathematical arguments. This direction explores a new research paradigm of "human researcher + AI tools + formal verification."

Contact

  •   qzyang@shnu.edu.cn
  •   College of Information, Mechanical and Electrical Engineering, SHNU