Reproduce “Adaptive Concept Bottleneck for Foundation Models Under Distribution Shifts” 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
Jihye Choi, Jayaram Raghuram, Yixuan Li, Somesh Jha
Researchers from the University of Wisconsin-Madison developed CONDA, a framework to enhance the interpretability and robustness of foundation model-based systems under distribution shifts. CONDA significantly boosts accuracy and provides adaptively aligned concept-based explanations using only unlabeled test data.
Key insights
- CBMs, even when built on powerful foundation models, can be less robust to distribution shifts than non-interpretable models if not adaptively adjusted.
- Different types of distribution shifts (low-level vs. concept-level) require distinct adaptive mechanisms, validating CONDA's multi-component design.
- Dynamically enriching the concept set (via RCB) is crucial for addressing 'incomplete concept set' failure modes, allowing for the discovery of new, relevant interpretive concepts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.