Researchers propose PeTeR, a novel data-free post-training framework designed to robustify pre-trained probabilistic circuits (PCs) against distribution shifts without requiring retraining from scratch. Standard likelihood-based PC learning is often vulnerable to overfitting and fragile generalization when facing data noise or small sample sizes.
- PeTeR mitigates these issues by applying distributionally-robust optimization that considers worst-case distributions within a Wasserstein ball of the empirical distribution.
- The method operates as a post-training step, allowing existing models to be robustified without the computational cost of full retraining.
- Empirical evaluations across multiple density estimation benchmarks show PeTeR effectively handles both random and adversarial perturbations.
- The approach achieves competitive or superior performance compared to data-dependent robust learning baselines.
This framework provides a practical solution for improving the reliability of probabilistic circuits in noisy or shifting environments without the overhead of training new models from scratch.