Основные понятия
Different divergences in variational inference yield varying estimates of uncertainty, leading to an ordering based on marginal variances.
Аннотация
The content delves into the analysis of different divergences in variational inference, focusing on factorized Gaussian approximations. It explores the impact of divergence choice on the estimation of uncertainty measures like variance, precision, and entropy. The structure is divided into sections discussing KL divergence, R´enyi divergence, score-based divergence, and an impossibility theorem for factorized VI. The implications of these analyses are then used to establish an ordering of divergences based on their performance in estimating uncertainty measures.
1. Introduction to Variational Inference
Problem: Approximating intractable distribution p by tractable q within family Q.
Importance of Variational Inference (VI) in Bayesian statistics and machine learning.
2. Divergence Definitions
Explanation of Kullback-Leibler (KL), R´enyi, and score-based divergences.
Comparison between reverse and forward KL divergences.
3. Analysis of Divergence Minimization
Detailed examination of minimizing KL and R´enyi divergences for factorized Gaussian VI.
Formulation and optimization using NQP for score-based divergences.
4. Ordering of Divergences
Definition: Dominance based on marginal variances estimation.
Propositions establishing relationships between precisions, entropies, and dominance.
Theorem presenting a comprehensive ordering of divergences for variational inference.
Статистика
"The α-divergence has also been studied in the context of VI (Li and Turner, 2016)."
"It has been widely observed that VI based on the reverse KL divergence tends to underestimate uncertainty."
"Proposition 3.1 (Precision matching). If q in eq. (4) has a diagonal covariance but p in eq. (3) does not..."
Цитаты
"The goal is to discover the best-matching distribution q within some larger parameterized family Q."
"VI often achieves a smaller error than Markov chain Monte Carlo methods with finite computational budgets."