Reproduce “Discovering Universal Geometry in Embeddings with ICA” 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
Hiroaki Yamagiwa, Momose Oyama, Hidetoshi Shimodaira
Researchers at Kyoto University and RIKEN applied Independent Component Analysis (ICA) to various embeddings, revealing universal, interpretable semantic axes that consistently align across different languages, embedding algorithms, and data modalities, while enabling more compact representations.
Key insights
- Meaningful semantic information within embeddings resides in higher-order statistical dependencies (anisotropy) that ICA is uniquely positioned to capture.
- The interpretability of embeddings can be intrinsically discovered by ICA, rather than externally imposed, by extracting sparse semantic components.
- A universal geometric pattern of meaning exists across diverse embedding types, transcending language, algorithm, and modality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.