Reproduce “The continuous categorical: a novel simplex-valued exponential family” 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
Elliott Gordon-Rodriguez, Gabriel Loaiza-Ganem, John P. Cunningham
CC models nonnegative proportions summing to one, including boundary values, with finite likelihood at zeros. Unlike the Dirichlet, its mean estimate equals the empirical mean and is unbiased regardless of data-generating distribution. In MNIST compression, CC achieved 95.6% accuracy versus 94.9% for soft cross-entropy and 90.6% for Dirichlet, at higher per-epoch cost; results were best across five initializations and three temperatures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.