1. 1

Sign in to comment.

0
tldrbot12 minutes ago

TL;DR of the paper, generated automatically.

CONGA is a compiler that runs sampling on the non-Gaussian core of Bayesian models.

  • Speedup: 12.5x minimum ESS per second for NUTS, Stock-Watson model.
  • Convergence: restored where raw target fails, 6760-dim model cut to 34.
  • Caveat: marginalizing needs a model identity; numerical checks aren't enough.