Sitemap
A list of all the posts and pages found on the site. For you robots out there, there is an XML version available for digesting as well.
Pages
Yeqi He’s Personal Website
Yeqi He
Posts
多层感知机
Published:
本文详细介绍了从感知机到多层感知机（MLP）的演变历程，分析了 XOR 问题对 AI 发展的影响及非线性激活函数的重要性。文章深入讲解了 MLP 的结构、常用激活函数（Sigmoid、Tanh、ReLU）及基于 PyTorch 的代码实现，并探讨了参数初始化、深层网络优势及“维度灾难”等核心问题。
线性回归
Published:
本文深入探讨了线性回归模型的基本原理，将其等价为单层神经网络进行解析。详细介绍了损失函数、代价函数、解析解及梯度下降优化算法的数学原理，并分别展示了基于 PyTorch 的从零代码实现与利用框架组件的简洁实现，为理解深度学习模型的训练过程奠定基础。
自动求导
Published:
本文深入探讨了深度学习中自动求导的核心机制。首先讲解了向量与矩阵求导的链式法则及具体推导示例；接着详细解析了计算图（Computational Graph）的原理、正反向传播的计算与内存复杂度；最后介绍了基于 PyTorch 的自动微分代码实现与常用操作，为理解深度学习框架的底层运算逻辑奠定基础。
矩阵计算
Published:
本文深入探讨了矩阵求导在深度学习中的意义，详细介绍了标量、向量、矩阵之间相互求导的规则和形状变化，并总结了分子布局下的求导公式，为理解反向传播算法奠定数学基础。
线性代数
Published:
本文总结了深度学习中常用的线性代数基础知识，包括向量点乘、范数、矩阵乘法、特征值分解等基本运算，并提供了对应的 PyTorch 代码实现，最后解答了关于 Tensor 数据类型转换的常见问题。
数据操作及数据预处理
Published:
本文介绍了深度学习中 N 维数组的概念，以及使用 PyTorch 和 Pandas 进行数据操作和预处理的基本方法，包括张量的创建、运算、广播机制以及 CSV 文件的读取和缺失值处理。
Git 快速入门及在线协同
Published:
本文介绍了 Git 的安装与 VSCode 集成配置，讲解了 Git 的基本语法，并详细说明了如何使用 Github 进行多人在线协同，包括贡献者协同（Pull Requests）和核心开发组内协同的流程。
Markdown 快速入门
Published:
本文介绍了如何配置 VSCode 以优化 Markdown 编写体验，包括推荐安装的扩展插件如 Markdown All In One、Markdown Emoji 等，以及如何利用这些工具提升文档编写效率。
publications
Checking the Pulse and Temperature of Higher Education
Published in Interdisciplinary Contest in Modeling (ICM), 2021
Many countries in the world are paying increasing attention to the health of the higher education system which is related to the level of national development. Therefore, we build models for the health and sustainability of the higher education system and evaluate the level of higher education in each country and give effective recommendations.
For Issue 1 and 2, we first collect ten important indexes that affect the evaluation of higher education systems in the United States, Canada, Switzerland, and China. They include, for example, the employment and unemployment rate of college students, and the proportion of the population with higher education. Then we conduct Analytic Hierarchy Process through the evaluation data of two experts and determine the hierarchical and weight relationship between the indexes. Finally, we calculate the final judgement result of each country through the Fuzzy Comprehensive Evaluation method. In 2019, the final judgement results of the United States, Canada, Switzerland and China are 0.7912, 0.7703, 0.7827, and 0.7545(see Figure 6).
For Issue 3 and 4, we take China, who has the lowest final judgement result obtained in Issue 1 and 2, as our research object. By analyzing the weights and scores of the indexes, we select the proportion of the population in higher education(PPHE) and the research quality index(RQI) as indexes that need improvement. We collect their historical data and conduct Time Series Analysis. Then we use the Damped Trend Model to make future predictions for them. We use the 2020 Swiss final judgement result as the threshold. When China’s final judgement result reaches 0.784237 which exceeds the threshold, we select the PPHE 42.56% and RQI 61.72 for the year as the result and regard them as the vision for China’s higher education system.
For Issue 5, we analyze the PPHE and RQI. Through the analysis of China’s national conditions, we decide to focus on adult higher education. We first predict the number of people that the new policy need to promote. Then we assess the strength of higher education of provinces in mainland China through the Analytic Hierarchy Process and get the number of people need to increase in each province. In the end, we analyze the 10-year policy arrangements for each province. We propose that universities divide the subjects into practical-based and research-based subjects. The classification of subjects not only makes application-oriented students more competitive after graduation, but also makes research-oriented students more capable of research.
For Issue 6 and 7, we use the model in Issue 2 to evaluate the theoretical results of the new policy and get the final judgement result of China’s higher education system in the next ten years. The new policy brings the final judgement result to 0.8030 in 2030, which is greater than the one without the new policy and the effect is very significant.Finally, we discuss the impact and difficulties during policy implementation from two aspects: ownership of management rights and impact of market economy.
Cite: Yeqi He, Penghui Xue and Hongming Liu. Checking the Pulse and Temperature of Higher Education[J]. Interdisciplinary Contest in Modeling (ICM), 2021.
Publisher | BibTeX
Few-Shot Generative Model Adaption via Identity Injection and Preservation
Published in Arxiv, Submitted to ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 2026
Training generative models with limited data presents severe challenges of mode collapse. A common approach is to adapt a large pretrained generative model upon a target domain with very few samples (fewer than 10), known as few-shot generative model adaptation. However, existing methods often suffer from forgetting source domain identity knowledge during adaptation, which degrades the quality of generated images in the target domain. To address this, we propose Identity Injection and Preservation (I2P), which leverages identity injection and consistency alignment to preserve the source identity knowledge. Specifically, we first introduce an identity injection module that integrates source domain identity knowledge into the target domain’s latent space, ensuring the generated images retain key identity knowledge of the source domain. Second, we design an identity substitution module, which includes a style-content decoupler and a reconstruction modulator, to further enhance source domain identity preservation. We enforce identity consistency constraints by aligning features from identity substitution, thereby preserving identity knowledge. Both quantitative and qualitative experiments show that our method achieves substantial improvements over state-of-the-art methods on multiple public datasets and 5 metrics.
Cite: Yeqi He, Liang Li, Jiehua Zhang, Yaoqi Sun, Xichun Sheng, Zhidong Zhao, and Chenggang Yan. Few-Shot Generative Model Adaption via Identity Injection and Preservation[J]. arXiv preprint arXiv:2603.22965, 2026.
Publisher | BibTeX
HAM: A Training-Free Style Transfer Approach via Heterogeneous Attention Modulation for Diffusion Models
Published in the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2026 (CVPR 2026)
Diffusion models have demonstrated remarkable performance in image generation, particularly within the domain of style transfer. Prevailing style transfer approaches typically leverage pre-trained diffusion models’ robust feature extraction capabilities alongside external modular control pathways to explicitly impose style guidance signals. However, these methods often fail to capture complex style reference or retain the identity of user-provided content images, thus falling into the trap of style-content balance. Thus, we propose a training-free style transfer approach via heterogeneous attention modulation (HAM) to protect identity information during image/text-guided style reference transfer, thereby addressing the style-content trade-off challenge. Specifically, we first introduces style noise initialization to initialize latent noise for diffusion. Then, during the diffusion process, it innovatively employs HAM for different attention mechanisms, including Global Attention Regulation (GAR) and Local Attention Transplantation (LAT), which better preserving the details of the content image while capturing complex style references. Our approach is validated through a series of qualitative and quantitative experiments, achieving state-of-the-art performance on multiple quantitative metrics.
Cite: Yeqi He, Liang Li, Zhiwen Yang, Xichun Sheng, Zhidong Zhao, and Chenggang Yan. HAM: A Training-Free Style Transfer Approach via Heterogeneous Attention Modulation for Diffusion Models[J]. arXiv preprint arXiv:2603.24043, 2026.
Publisher | BibTeX