Core Concepts
COLEP proposes a certifiably robust learning-reasoning conformal prediction framework via probabilistic circuits, achieving higher prediction coverage and accuracy.
Abstract
The content introduces COLEP as a framework for conformal prediction using probabilistic circuits. It discusses the challenges of adversarial perturbations in traditional models and presents the theoretical foundation and certification process for COLEP. The reasoning component with PCs is detailed, along with the construction of knowledge rules. The effectiveness of the reasoning component is analyzed, showing its impact on improving prediction accuracy and coverage. Experiments on various datasets demonstrate the superiority of COLEP over traditional models in terms of certified coverage and prediction accuracy under bounded perturbations.
Structure:
Introduction to COLEP
Theoretical Foundation and Certification Process
Reasoning Component with Probabilistic Circuits
Effectiveness Analysis of Reasoning Component
Comparison with Traditional Models
Experimental Results
Stats
ˆπCOLEP achieves higher certified coverage.
Empirical validation on GTSRB, CIFAR-10, AwA2 datasets.