Reproduce “Concept Bottleneck Model with Additional Unsupervised Concepts” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yoshihide Sawada, Keigo Nakamura
Researchers from Aisin's Tokyo Research Center developed the Concept Bottleneck Model with Additional Unsupervised Concepts (CBM-AUC), which integrates supervised human-interpretable concepts with unsupervised learned representations. This model achieved a classification error of 0.177 on the CUB-200-2011 dataset, outperforming the standard CBM (0.199 error), and demonstrated scalability to large images in autonomous driving scenarios while maintaining interpretability.
Key insights
- Combining explicit human-defined concepts with implicitly learned, unsupervised representations can enhance model performance without sacrificing interpretability.
- Computational efficiency improvements (discriminator-based learning, weight-shared intermediate network) are crucial for scaling interpretable models to real-world, high-resolution image datasets.
- Unsupervised concepts provide complementary information to supervised ones, making the model more robust to situations with limited supervised concept labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.