Computational study demonstrates competitive accuracy and transparent reasoning across benchmark datasets, suggesting a viable interpretable alternative to black-box neural networks.
Key Points
To develop an inherently transparent neural network architecture that combines attribute prediction with probabilistic logical reasoning for accurate, interpretable classification.
Designed Neural Probabilistic Circuits (NPCs) comprising an attribute recognition module paired with a probabilistic circuit predictor for compositional logical reasoning.
Implemented a three-stage training pipeline consisting of attribute recognition training, circuit construction, and joint model optimization.
Evaluated performance, theoretical error upper bounds, and generated most-probable and counterfactual explanations across four benchmark datasets.
Theoretical analysis proved that an NPC's overall error is upper-bounded by a linear combination of the errors from its individual modules.
Empirical evaluations across four benchmark datasets demonstrated that NPCs achieve predictive accuracy competitive with traditional end-to-end black-box deep neural networks.
NPCs provided faithful interpretability by generating verifiable logical explanations, including both most probable explanations and counterfactual scenarios.