Research paper explores computation in neural superposition

A new research paper titled "The Ball and the Box: Two Geometries of Computation in Superposition" explores how neural representations can encode more features than their dimensions, a phenomenon known as superposition. The study focuses on the dimensional requirements for computing Boolean gates from these representations. It derives sharp dimension thresholds under two error criteria for a single threshold layer with a Gaussian random dictionary and sparse Boolean inputs. The paper highlights that a vanishing expected error count may necessitate more dimensions than ensuring correctness for every output with high probability, attributing this gap to shared reads where rare realizations can cause multiple errors simultaneously. AI

RANK_REASON The item is a research paper submitted to arXiv. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• arXivLabs
• Bibliographic Explorer
• Boolean
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gaussian function
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• Litmaps
• ScienceCast
• scite Smart Citations
• The Ball and the Box: Two Geometries of Computation in Superposition

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →