Learning Probabilistic Logic Models over Structured and Unstructured Data
Keywords: probabilistic logic models, relational dependency networks, neurosymbolic learning, knowledge-based learning
Abstract: Effective decision-making in high-stakes domains necessitates reconciling information from structured and unstructured data with incomplete and imprecise background knowledge. Relational Dependency Networks are a popular class of probabilistic logic models that support efficient reasoning over structured data and symbolic domain knowledge but struggle to accommodate unstructured data such as images and text. On the other hand, neural networks excel at extracting patterns from unstructured data but are not amenable to reasoning. We propose Deep Relational Dependency Networks which combine Relational Dependency Networks with neural networks to reason effectively about multimodal data and symbolic domain knowledge. Experiments on scene classification tasks with noisy and limited data indicate that this approach yields more accurate yet interpretable models.
Submission Number: 23
Loading