- Label-Free Idea Bottleneck Fashions
Authors: Tuomas Oikarinen, Subhro Das, Lam M. Nguyen, Tsui-Wei Weng
Summary: Idea bottleneck fashions (CBM) are a well-liked manner of making extra interpretable neural networks by having hidden layer neurons correspond to human-understandable ideas. Nevertheless, current CBMs and their variants have two essential limitations: first, they should accumulate labeled knowledge for every of the predefined ideas, which is time consuming and labor intensive; second, the accuracy of a CBM is commonly considerably decrease than that of a regular neural community, particularly on extra complicated datasets. This poor efficiency creates a barrier for adopting CBMs in sensible actual world functions. Motivated by these challenges, we suggest Label-free CBM which is a novel framework to remodel any neural community into an interpretable CBM with out labeled idea knowledge, whereas retaining a excessive accuracy. Our Label-free CBM has many benefits, it’s: scalable — we current the primary CBM scaled to ImageNet, environment friendly — making a CBM takes only some hours even for very giant datasets, and automatic — coaching it for a brand new dataset requires minimal human effort. Our code is accessible at https://github.com/Trustworthy-ML-Lab/Label-free-CBM. Lastly, in Appendix B we conduct a big scale consumer analysis of the interpretability of our methodology.
