Reproduce “Learning explanations that are hard to vary” 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
Giambattista Parascandolo, Alexander Neitz, Antonio Orvieto, Luigi Gresele, Bernhard Schölkopf
This paper presents a method to enhance out-of-distribution generalization in deep learning by prioritizing explanations that are "hard to vary." It introduces the "AND-mask" algorithm, which filters inconsistent gradient signals across examples or environments, achieving superior robustness against spurious correlations and label noise across various tasks.
Key insights
- Standard gradient averaging can be viewed as a "logical OR" operation, which can lead to learning inconsistent or spurious patterns.
- Robust, generalizable explanations should be "hard to vary" across different data contexts, meaning their underlying mechanisms remain consistent.
- Invariances can be discovered and enforced even without explicit environment labels, by treating individual examples or mini-batches as distinct environments for consistency checks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.