Expressive Power of Deep Homomorphism Networks over Relational Databases
Researchers have introduced Deep Homomorphism Networks (DHNs) as a powerful architecture for learning from relational databases, drawing parallels to fragments of SQL. Their study connects DHNs with various extensions of first-order logic, including those with counting and ratio quantifiers. These findings also shed light on the decidability of static analysis problems for DHNs and are supported by experimental results showing performance differences that align with their expressive power. AI
IMPACT Introduces a new model architecture with theoretical connections to SQL, potentially improving database learning tasks.