Reproduce “Removing Spurious Concepts from Neural Network Representations via Joint Subspace Estimation” 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
Floris Holstege, Bram Wouters, Noud van Giersbergen, Cees Diks
Joint Subspace Estimation (JSE) precisely removes spurious concepts from neural network representations by jointly estimating orthogonal subspaces for spurious and main-task features. This method preserves main-task performance and enhances out-of-distribution generalization, overcoming the issue of existing techniques inadvertently deleting relevant main-task information.
Key insights
- The primary limitation of existing concept-removal methods stems from their inability to distinguish between truly spurious features and main-task features spuriously correlated with them.
- Jointly estimating and enforcing orthogonality between spurious and main-task concept subspaces during the removal process is critical for preventing the unintended deletion of main-task information.
- Statistical tests for both informativeness and dominant predictiveness allow for a data-driven, precise determination of concept subspace dimensions, ensuring accurate and non-destructive concept removal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.