This page summarizes my research interests, publications, projects, research experience, and other academic resources.

本页汇总了我的研究方向、论文、项目、研究经历以及其他与学术研究相关的内容。

Research Interests

研究方向

Currently focusing on industry research:

AI Applications: AI + Foundational Software Innovation

During my PhD, I mainly focused on:

Machine Learning: Continual Learning
Trustworthy AI: Continual Unlearning, Machine Unlearning, Explainable AI

目前关注业界研究：

AI 应用：AI+基础软件创新

博士期间主要关注：

机器学习（Machine Learning）：持续学习（Continual Learning）
可信人工智能（Trustworthy AI）：持续遗忘（Continual Unlearning）、机器遗忘（Machine Unlearning）、人工智能可解释性（Explainable AI）

Publications

发表论文

AdaHAT: Adaptive Hard Attention to the Task in Task-Incremental Learning

Pengxiang Wang, Hongbo Bo, Jun Hong, Weiru Liu, and Kedian Mu.
Proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases (ECML PKDD), 2024, pp. 143–160.

Pengxiang Wang、Hongbo Bo、Jun Hong、Weiru Liu、Kedian Mu
发表于会议 ECML PKDD 2024（European Conference on Machine Learning and Knowledge Discovery in Databases）

DOI arXiv Slides Poster Code

DOI arXiv 幻灯片海报代码

Fine-Grained Neuron Importance Guided Architecture-Based Continual Learning

Pengxiang Wang, Rongyu Zhu, Junjie Liu, and Kedian Mu
Accepted by the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD), 2026. Will be published in the conference proceedings in August.

Pengxiang Wang、Rongyu Zhu、Junjie Liu、Kedian Mu
已被会议 ECML PKDD 2026（European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases）接受，8月份出版

Slides Poster Code

幻灯片海报代码

Unlearn to Learn Better: Continual Unlearning via Capacity Recycling

Pengxiang Wang, Shaobo Zhang, Hongbo Bo, and Kedian Mu
Accepted by the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD), 2026. Will be published in the conference proceedings in August.

Pengxiang Wang、Shaobo Zhang、Hongbo Bo、Kedian Mu
已被会议 ECML PKDD 2026（European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases）接受， 8月份出版

Slides Poster Code

幻灯片海报代码

Predicting Student Performance Using Weblogs

Bo Wu, Shaojie Qu, Yin Ni, Yemin Zhou, Pengxiang Wang, Qiwen Li
2019 14th International Conference on Computer Science & Education (ICCSE), 2019, pp. 616–621.

Bo Wu、Shaojie Qu、Yin Ni、Yemin Zhou、Pengxiang Wang、Qiwen Li
发表于会议 ICCSE（14th International Conference on Computer Science & Education）2019

DOI

Open-Source Projects

开源项目

Continual Learning Arena (CLArena)

持续学习竞技场（CLArena）

An open-source Python package for continual learning research, published on PyPI
Independently designed, developed, tested, documented, and maintained by myself

面向持续学习领域研究的开源 Python 软件包，发布于 PyPI
由我独立完成设计、开发、测试、文档撰写与维护

Project Homepage PyPI GitHub API Reference

项目主页 PyPI GitHub API 文档

Research Experience

研究经历

In July 2026, I received my PhD in Applied Mathematics from the School of Mathematical Sciences at Peking University and subsequently moved into industry, bringing my doctoral research journey to a close.

2026 年 7 月，我从北京大学数学科学学院应用数学专业博士毕业，随后进入业界工作，我的博士研究至此告一段落。

My research interests have always centered on artificial intelligence. When I was an undergraduate, I majored in Information and Computing Science (belongs to mathematics major). Alongside my mathematical training, I also maintained a strong interest in computer science. Starting in my third year, I systematically studied machine learning and deep learning through coursework and self-directed learning. I participated in an machine learning project in education led by my data structures instructor, published my first paper, and developed a machine learning-based photovoltaic power generation forecasting model for my bachelor’s dissertation. This marked the beginning of my research.

我的研究兴趣一直集中在人工智能领域。本科阶段，我主修信息与计算科学专业（数学专业下的本科小专业），在数理训练之外也一直对计算机抱有浓厚兴趣。大三起，我通过课程与自学系统接触机器学习、深度学习，参与了本科数据结构老师的面向教育场景的机器学习应用项目，发表了第一篇论文，并在本科毕业论文中完成了一个基于机器学习的光伏发电功率预测模型——这是我走上研究道路的开始。

During my PhD, under the supervision of Dr. Kedian Mu, I explored several active research areas at the time, including continual learning, meta-learning, and multi-task learning. I eventually focused on architecture-based continual learning, which led to my first major research project. During this period, I spent one year at the University of Bristol as a visiting PhD student. Under the supervision of Prof. Weiru Liu and Prof. Jun Hong, I developed this work into a publication while also becoming acquainted with trustworthy AI and explainable AI. After returning to China, I further integrated the ideas from these fields with continual learning and completed two additional projects. All three papers were published or accepted at the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD). I ultimately completed my PhD thesis, titled Continual Learning and Unlearning Algorithms in Deep Neural Networks. I’d like to express my sincere gratitude to my supervisors and collaborators—Hongbo Bo, Shaobo Zhang, Rongyu Zhu, and Junjie Liu —for their support and contributions to this research. For details of these papers, please refer to the Publications section above.

博士阶段，我在导师牟克典副教授的指导下，先后探索了持续学习、元学习和多任务学习等在当时比较火热的研究方向，最终将研究重点放在基于网络结构的持续学习上，并由此形成第一项工作的原型。其间，我来到英国进行了为期一年的博士联合培养，在 Weiru Liu 教授和 Jun Hong 教授的指导下将该工作整理并发表，同时接触到了可信人工智能、人工智能可解释性等领域。回国后，我进一步将这些思路与持续学习相结合，完成了另外两项工作。三篇论文均发表或录用于欧洲机器学习与数据挖掘会议（ECML PKDD）。最终，我以《深度神经网络中的持续学习与遗忘算法研究》为题完成博士学位论文。在此感谢导师们与合作者 Hongbo Bo 师兄、Shaobo Zhang 师弟、Rongyu Zhu 师弟和 Junjie Liu 师弟对我的支持和对这些研究工作的贡献。关于这些论文的详细信息，请参见上方的发表论文部分。

One distinctive outcome of my PhD was an open-source software for artificial intelligence research: the Python package Continual Learning Arena (CLArena), entirely designed, developed, tested, documented, and maintained by myself. It supports a range of machine learning paradigms, including continual learning, continual unlearning, multi-task learning, and single-task learning, coming with comprehensive documentation and API references. The motivation for developing CLArena arose from practical problems that I repeatedly encountered in my research. Conducting experiments directly on top of different authors’ codebases often involved high reproduction costs, while comparing new methods required considerable effort to adapt implementations with widely differing structures. Some codebases were even not sufficiently robust or complete to reliably reproduce the experimental results reported in the corresponding papers. In the third year of my PhD, I decided to build my own continual learning experimentation platform from scratch and treat the process as an opportunity to systematically strengthen my programming skills. At a time when large language models were not yet widely used for programming, completing even the first version of CLArena took several months, but I persevered. I continued to iterate on the project, eventually developing it into a fully featured continual learning framework with extensive interfaces for customization, together with implementations of numerous continual learning algorithms, datasets, network architectures, etc. CLArena supported my three major research projects. A unified codebase allowed me to understand every detail of my experiments and gave me greater confidence in the reliability of my results at critical stages of my PhD. Today, as large language models and vibe coding become increasingly popular, a project written in this old-fashioned, manual way may feel like a relic of another era. For me, however, it was an invaluable training experience. It enabled me to accumulate substantial programming and software development experience during my PhD, and it also helped me secure an industry role in foundational software that I genuinely wanted. Looking back, I believe it was all worthwhile. For more information about CLArena, please refer to the Open-Source Projects section above.

在博士阶段，我的一项较有特色的产出是一款面向人工智能领域研究的开源软件——Python 软件包 Continual Learning Arena（CLArena），由我独立完成设计、开发、测试、文档撰写与维护。该软件包支持持续学习、持续遗忘、多任务学习和单任务学习等多种机器学习范式，配有完善的英文文档与 API 文档。CLArena 的起点，是我在科研中反复遇到的一些实际问题：直接基于不同作者的代码库开展实验，不仅复现成本较高，在比较新方法时也需要投入大量时间适配彼此差异很大的实现；有些代码甚至本身的质量和完整性也不够理想，难以可靠支撑论文中的实验结果。博士三年级时，我决定从头搭建一套属于自己的持续学习实验平台，也把它当作一次系统锻炼编程能力的机会。在大模型尚未广泛参与编程的那个时候，仅完成 CLArena 的初版就花费了数月，但我还是坚持了下来。此后，我持续迭代这个项目，将其发展为一套功能完整、提供丰富自定义接口的持续学习训练框架，并在其上实现了多种持续学习算法、数据集和网络结构。CLArena 支撑了我的三项主要研究工作：统一的代码体系让我能够掌握实验的每一处细节，也让我在读博的关键节点上对结果的可靠性更有信心。在大模型普及、vibe coding 流行的今天，这样古法编程出来的项目成了时代的眼泪，但对我来说，这是一次非常宝贵的训练经历。我由此在博士阶段积累了丰富的编程与软件开发经验，这段经历也帮助我找到了心仪的基础软件领域的业界工作。回过头来看，我觉得这一切都是值得的。关于 CLArena 的更多信息，请参见上方的开源项目部分。

Next Chapter

下一阶段

For the foreseeable future, my research in continual learning will be on pause. I will shift my focus toward industry research in AI applications and foundational software innovation, with less emphasis on academic research and greater attention to real-world deployment and open-source projects. Please stay tuned! :)

未来一段时间，我在持续学习领域的研究将告一段落。接下来，我将转向人工智能应用与基础软件创新方向的业界研究，可能较少涉及学术研究，而是更多地关注技术落地与开源项目。敬请期待！:)

Other Academic Resources

其他学术资源

During my PhD, I also accumulated a collection of tutorials, study notes, paper notes, code examples, and research experience on this website. You are welcome to explore them.

博士期间，我也在这个网站中积累了不少教程、学习笔记、论文笔记、代码示例与科研经验，欢迎查看。

I remain open to academic and technical discussions. If you would like to discuss AI-related topics or have ideas worth exploring together, please feel free to contact me through any of the social platforms at the top of this page.

我对学术与技术交流保持开放。如果你希望讨论 AI 相关的问题，或有任何值得一起探索的想法，欢迎通过页面顶部的社交平台与我联系。