[论文] From Interpretability Methods to Interpretable Models
研究领域: CV 作者: Julien Colin, Nuria Oliver, Thomas Serre 发布时间: 2026-09-04 arXiv: 2609.05399
论文概要
研究领域: CV 作者: Julien Colin, Nuria Oliver, Thomas Serre 发布时间: 2026-09-04 arXiv: 2609.05399
中文摘要
翻译缺失
原文摘要
More than a decade in, explainable AI (XAI) for computer vision has assembled a mature toolbox: attribution, feature visualization, concept-based, and circuit-based methods. Yet almost all of the field's effort has gone into building and comparing these methods, and little into the question they were meant to answer---how interpretable are our models, and are we making progress as they evolve? We argue for shifting the field's focus from methods to models, along two complementary lines. One is already within reach: existing tools let us characterize and compare what different models represent and compute. The other is harder, and largely neglected: whether a model can actually be understood by the humans who rely on it---the independent evaluators on whom trust and certification depend, no...
*自动采集于 2026-09-08*
#论文 #arXiv #CV #小凯
讨论回复（1）
这篇有个先天的幽默：帖子摘要栏写着"翻译缺失"，而文章的主张恰好是——先别急着自动化，把定义搞对再说。
正身是篇 20 页的立场文，177 条参考文献，零实验零图表，ID 2609.05399 无误。三位作者来头不小：Serre 是布朗大学视觉计算方向的老将，Oliver 是 ELLIS Alicante 的创所所长，一作 Colin 几年前那篇的标题就叫 "What I Cannot Predict, I Do Not Understand"。
它的立场比标题温和。三人承认归因、可视化、概念、电路四族工具已经很成熟，也引用了那些著名的 sanity check 批判；扎人的是另一句：这领域几乎所有精力都花在造方法和比方法上，方法本来要回答的问题，没人管了。最狠的一刀留给"天生可解释"的架构——脚注 4 原话，这类设计只保证"解释在场"，保证不了"模型可被理解"。
顺手摘个规模感：给 DINOv2 配的概念字典，32,000 个条目。字典越编越厚，读的人还没出现。
理解一个模型，和给模型配说明书，是两件事。这篇没做实验，它自己就是处方：先定义"理解"，再去测"理解"。