[Paper Review] Belief Optimization for Binary Networks: A Stable Alternative to Loopy Belief Propagation

Max Welling, Yee Whye Teh|arXiv (Cornell University)|Jan 10, 2013

Bayesian Modeling and Causal InferenceComputer Science16 references102 citations

TL;DR

This paper proposes belief optimization, a stable inference algorithm for binary undirected graphical models that minimizes the Bethe free energy via alternating updates of pairwise and marginal probabilities. Unlike loopy belief propagation, it guarantees convergence to a local minimum and outperforms BP in non-convergent cases, making it ideal for learning graphical models from data.

ABSTRACT

We present a novel inference algorithm for arbitrary, binary, undirected graphs. Unlike loopy belief propagation, which iterates fixed point equations, we directly descend on the Bethe free energy. The algorithm consists of two phases, first we update the pairwise probabilities, given the marginal probabilities at each unit,using an analytic expression. Next, we update the marginal probabilities, given the pairwise probabilities by following the negative gradient of the Bethe free energy. Both steps are guaranteed to decrease the Bethe free energy, and since it is lower bounded, the algorithm is guaranteed to converge to a local minimum. We also show that the Bethe free energy is equal to the TAP free energy up to second order in the weights. In experiments we confirm that when belief propagation converges it usually finds identical solutions as our belief optimization method. However, in cases where belief propagation fails to converge, belief optimization continues to converge to reasonable beliefs. The stable nature of belief optimization makes it ideally suited for learning graphical models from data.

Motivation & Objective

• To address the instability and non-convergence issues of loopy belief propagation in binary undirected graphical models.
• To develop a provably convergent inference algorithm that minimizes the Bethe free energy.
• To provide a robust alternative suitable for learning graphical models from data.
• To establish a connection between the Bethe free energy and the TAP free energy up to second order in weights.

Proposed method

• The algorithm performs two alternating phases: updating pairwise probabilities given fixed marginals using an analytic expression.
• It then updates marginal probabilities by following the negative gradient of the Bethe free energy.
• Both update steps are mathematically guaranteed to decrease the Bethe free energy.
• The algorithm's convergence is ensured because the Bethe free energy is bounded from below.
• The method leverages the equivalence between the Bethe free energy and the TAP free energy up to second-order terms in the weights.
• The approach avoids fixed-point iteration, instead using gradient descent on the free energy landscape.

Experimental results

Research questions

• RQ1 Can a stable inference algorithm be developed for binary undirected graphical models that avoids the convergence issues of loopy belief propagation?
• RQ2 Does minimizing the Bethe free energy via gradient-based updates lead to reliable and convergent belief estimation?
• RQ3 How does the proposed belief optimization method compare to loopy belief propagation in terms of convergence and solution quality?
• RQ4 What is the relationship between the Bethe free energy and the TAP free energy in binary networks?

Key findings

• Belief optimization guarantees convergence to a local minimum of the Bethe free energy due to its monotonic decrease and boundedness.
• When loopy belief propagation converges, belief optimization typically finds the same solution, confirming consistency.
• In cases where loopy belief propagation fails to converge, belief optimization continues to converge to meaningful, stable beliefs.
• The Bethe free energy is equivalent to the TAP free energy up to second order in the weights, establishing a theoretical link between the two frameworks.
• The algorithm's stability makes it particularly suitable for learning graphical models from data, where convergence is critical.