Abstract: Undirected graphical models are compact representations of joint probability distributions over random variables. To solve inference tasks of interest, graphical models of arbitrary topology can be trained using empirical risk minimization. However, to solve inference tasks that were not seen during training, these models (EGMs) often need to be re-trained. Instead, we propose an inference-agnostic adversarial training framework which produces an infinitely-large ensemble of graphical models (AGMs). The ensemble is optimized to generate data within the GAN framework, and inference is performed using a finite subset of these models. AGMs perform comparably with EGMs on inference tasks that the latter were specifically optimized for. Most importantly, AGMs show significantly better generalization to unseen inference tasks compared to EGMs, as well as deep neural architectures like GibbsNet and VAEAC which allow arbitrary conditioning. Finally, AGMs allow fast data sampling, competitive with Gibbs sampling from EGMs.
Introduction
Probabilistic graphical models (PGMs) are powerful tools for representing joint distributions, yet they face significant training challenges. Traditional Maximum Likelihood training is often computationally intractable due to the difficulty of calculating the partition function. To circumvent this, researchers frequently turn to Empirical Risk Minimization (ERM), creating models (EGMs) optimized for specific inference tasks. However, these "specialist" models often fail to generalize to tasks not encountered during training.
This research introduces Adversarially-trained Graphical Models (AGMs), a framework designed to create "generalist" models capable of handling arbitrary conditioning at test time without re-training.
Methods
The core of the AGM framework is the transition from learning a single set of parameters to learning an infinite ensemble of discrete undirected graphical models.
- The Infinite Ensemble: A neural network "Learner" ($L_\theta$) maps a latent noise vector $z$ to UGM parameters $\Psi$. This process defines an infinite pool of potential "local expert" models.
- Adversarial Training: Utilizing the Wasserstein GAN with Gradient Penalty (WGAN-GP) framework, the generator outputs UGM parameters rather than raw data. Belief Propagation (BP) is then used to produce marginal probabilities, which are evaluated by a discriminator against real data.
- Log-Linear Pooling: During inference, $M$ models are sampled from the ensemble. Their individual outputs are aggregated via log-linear pooling, which helps mitigate errors from any single model's mis-specification.
Key Results
The framework was evaluated on UCI binary datasets, binarized images (MNIST, Caltech-101), and discrete/RGB images (Stanford background, SVHN). AGMs demonstrate remarkable robustness across various scenarios:
- Performance on Known Tasks: AGMs perform nearly as well as task-specific EGMs on the queries those models were explicitly trained for, trailing by only ~2% on average.
- Superior Generalization: The primary advantage of AGMs is their ability to handle "unseen" tasks. When tested on arbitrary conditioning (e.g., predicting different image segments), AGMs significantly outperform both EGMs and advanced deep architectures like GibbsNet and VAEAC.
- Efficient Sampling: Unlike traditional UGMs that require slow, iterative Gibbs sampling, AGMs allow for "one-shot" data generation.
Conclusion and Future Work
While AGMs offer significant generalization benefits, they currently rely on fixed graph topologies. Future work may involve learning the graph structure dynamically. Additionally, while parallelizable, the computational cost of running Belief Propagation across a large ensemble is higher than standard neural network forward passes. Finally, extending the framework to natively support continuous variables remains an area for further exploration.
Summary updated