This website requires JavaScript.

Interactive Concept Bottleneck Models

Kushal ChauhanRishabh TiwariJan FreybergPradeep ShenoyKrishnamurthy Dvijotham
Dec 2022
摘要
Concept bottleneck models (CBMs) (Koh et al. 2020) are interpretable neuralnetworks that first predict labels for human-interpretable concepts relevant tothe prediction task, and then predict the final label based on the conceptlabel predictions.We extend CBMs to interactive prediction settings where themodel can query a human collaborator for the label to some concepts. We developan interaction policy that, at prediction time, chooses which concepts torequest a label for so as to maximally improve the final prediction. Wedemonstrate thata simple policy combining concept prediction uncertainty andinfluence of the concept on the final prediction achieves strong performanceand outperforms a static approach proposed in Koh et al. (2020) as well asactive feature acquisition methods proposed in the literature. We show that theinteractiveCBM can achieve accuracy gains of 5-10% with only 5 interactionsover competitive baselines on the Caltech-UCSDBirds, CheXpert and OAI datasets.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?
0
被引用
笔记
问答