Article Source

• PFNs; Use neural networks for 100x faster Bayesian predictions

PFNs; Use neural networks for 100x faster Bayesian predictions

• Title: Prior-data Fitted Networks (PFNs): Use neural networks for 100x faster Bayesian predictions
• Paper

Abstract

Bayesian methods can be expensive and complicated to approximate with e.g. MCMC or VI. PFNs are a new, cheap and simple method to accurately approximate Bayesian predictions. I will explain how to build a PFN out of a Transformer by meta-learning on artificial data. I present the results from our paper that introduces PFNs, in which PFNs beat VI and MCMC for some standard tasks, and a to-be-released follow-up work on Tabular classification, where we show that a simple PFN can replace a full AutoML tool in some scenarios.