Enhancing Interpretable Image Classification Through LLM Agents and Conditional Concept Bottleneck Models
Yiwen Jiang, Deval Mehta, Wei Feng, Zongyuan Ge
Monash University, Melbourne, Australia | AIM for Health Lab
Architectural comparison of CBMs and CoCoBMs
🚀 研究亮点

我们提出了条件概念瓶颈模型(CoCoBMs)和概念代理,通过动态调整概念库和类别特定评分机制,在6个数据集上实现了:

+6%
分类准确率提升
+30%
可解释性评估提升
动态
概念库优化
🔥 创新点
Concept Agent workflow
阅读完整论文