Notes & essays笔记与随笔
Working notes on evaluation, interpretability, and what makes biological AI results believable.一些工作笔记：评估、可解释性，以及什么样的生物 AI 结果才可信。
2026
- How Should We Design a Structured Null?如何设计结构化零假设？ Random label shuffles are too weak. Notes on building null distributions that preserve biological structure while destroying the signal you claim to detect.随机打乱标签往往太弱。本文讨论如何在保留生物结构的同时，破坏我们声称检测到的信号。
- Interpreting Protein Language Model Features如何解释蛋白质语言模型特征 What sparse autoencoders give us over raw neurons when analyzing PLM embeddings — and what they still cannot tell us.与直接分析单个神经元相比，稀疏自编码器能为 PLM 嵌入分析带来什么，又有哪些问题仍然无法回答。
- What Makes a Biological AI Result Convincing?什么样的生物 AI 结果才可信？ A personal checklist: external validation, calibration, leakage audits, and claim boundaries — the difference between a demo and a tool.我的检查清单：外部验证、校准、数据泄漏审计和论断边界——这些决定了一个模型只是演示，还是能真正使用的工具。
No posts match your search.没有匹配的文章。