Learning to Condition (L2C)
A scalable neural heuristic framework for accelerating Most Probable Explanation (MPE) inference in probabilistic graphical models.
Learning to Condition (L2C) is a scalable, data-driven framework for accelerating Most Probable Explanation (MPE) inference in Probabilistic Graphical Models (PGMs).
MPE inference—finding the most likely assignment to unobserved variables given evidence—is fundamentally NP-hard and computationally intractable in high-treewidth models. L2C trains a neural network to score variable-value assignments based on their utility for conditioning, given observed evidence, substantially reducing search spaces while maintaining or improving solution quality.
Links
- GitHub repository:
https://github.com/brijml/L2C - Paper: Learning to Condition: A Neural Heuristic for Scalable MPE Inference (NeurIPS 2025)
Problem Setting & Motivation
- Task (MPE Inference): Given a probabilistic graphical model defined over random variables $X$ and evidence $e$, find an assignment $x^*$ that maximizes the joint probability: \(x^* = \arg\max_{x} P(x \mid e)\)
- Challenge: In high-treewidth models, exact inference algorithms like variable elimination or junction tree require exponential time and memory in the treewidth.
- Conditioning Approach: Assigning values to a subset of variables simplifies the remaining problem by removing edges and reducing graph complexity. However, selecting which variables to condition on and which values to assign is a critical combinatorial challenge.
Key Ideas
- Neural Conditioning Heuristic: L2C trains a deep neural network that evaluates candidate variable-value assignments based on how effectively they reduce downstream search complexity without sacrificing solution quality.
- Scalable Data Generation: Extracts supervisory signals directly from the search traces of existing exact and approximate MPE solvers, circumventing the need for intractable ground-truth solutions during training data collection.
- Flexible Search Integration:
- Pre-conditioning: Applies the learned heuristic to condition high-impact variables prior to invoking exact inference solvers.
- Branch-and-Bound Guidance: Serves as a dynamic variable and value ordering policy within tree search algorithms.
- Empirical Scalability: Demonstrates significant reduction in search runtime and explored tree size across benchmark graphical models with challenging cyclic topologies.
Citation
If you use L2C in your research, please cite:
@inproceedings{malhotra2025learning,
author = {Malhotra, Brij and Arya, Shivvrat and Rahman, Tahrima and Gogate, Vibhav},
title = {Learning to Condition: A Neural Heuristic for Scalable MPE Inference},
booktitle = {Advances in Neural Information Processing Systems},
year = {2025},
volume = {38},
pages = {121270--121316},
publisher = {Curran Associates, Inc.}
}