我是 Wendy，一名深度学习学生，关注医学证据、生成模型与模型选择保留下来的视觉细节之间的关系。这里记录实验、笔记，以及还没有被解决的问题。I’m Wendy, a deep learning student tracing the space between medical evidence, generative models, and the visual details a model decides to keep. This is where experiments, notes, and unfinished questions stay visible.
一组小而清晰的研究方向：可比较、可控制，也能让结论经得起数据检验。A compact set of research directions designed for clear comparisons, small controlled experiments, and conclusions that survive contact with the data.
测试拓扑感知损失能否帮助模型保留普通重叠指标容易掩盖的细薄、连通结构。Testing whether topology-aware losses help models keep the thin, connected structures that ordinary overlap scores can hide.
从感受野、patch、注意力和曲线背后的归纳偏置出发，观察 CNN 与 ViT 的差异。A visual study of CNN and ViT: receptive fields, patches, attention, and the inductive biases hiding behind the curves.
通过代码理解 DDPM、DDIM 与 flow matching，再追问什么让合成眼底图像在结构上可信。Reading DDPM, DDIM, and flow matching through code, then asking what makes a synthetic fundus image structurally believable.
最好的解释通常来自一个小实验：只改变一件事，观察表示，再写下究竟发生了什么。The best explanation is usually a small experiment: change one thing, inspect the representation, and write down what moved.
特征图、attention rollout、去噪轨迹，以及模型信心开始变脆弱的地方。Feature maps, attention rollout, denoising trajectories, and the places where a model’s confidence becomes fragile.
学习率、优化器、归一化、损失、patch size：先做受控网格，再谈更大的理论。Learning rate, optimizer, normalization, loss, patch size: a controlled grid before a grand theory.
从公式写到张量形状，再写到实际后果，不跳过中间那些最容易卡住的部分。Notes that move from equation to tensor shape to practical consequence, without skipping the awkward middle.
关于重叠指标、连通性，以及“基本正确”和临床有用的结构之间的差别。On overlap metrics, connectivity, and the difference between “mostly right” and clinically useful structure.
从噪声出发、预测噪声，真正看懂状态如何更新，而不是只记住图。Starting from noise, predicting noise, and learning to see the state update rather than memorizing the diagram.
持续整理 locality、translation equivariance，以及 Transformer 不会自动拥有的视觉先验。A running glossary for locality, translation equivariance, and the visual priors a Transformer does not receive for free.
一个人、一个清晰对比，以及足够深入的分析，去弄清究竟改变了什么。One person, one clear comparison, enough depth to understand what actually changed.
我是北京大学药学院大二学生，正在把对生命科学的直觉，转译成可以运行、比较和解释的代码。现在最想弄清楚的，是模型如何从医学图像里学会看见结构。I’m a second-year student at Peking University’s School of Pharmaceutical Sciences, translating an intuition for life science into code that can run, compare, and explain. Right now I’m trying to understand how models learn to see structure in medical images.
我不想把药学和编程看成两条互相排斥的路；这个网站记录它们相遇的过程。I don’t see pharmacy and programming as separate paths. This site records what happens when they meet.
这里放蓝色笔触、结构图，以及让困难想法留在记忆里的视觉语言。A side room for blue brushstrokes, diagrams, and the visual language that makes a hard idea stay in memory.
用图、注释和实验，为模型隐藏的假设赋予形状。Diagrams, annotations, and experiments that give shape to a model’s hidden assumptions.
这里的每一页，都应该让一个问题比之前更清晰。Every page here should leave one question sharper than it found it.