Lifted Relational Neural Networks

We propose a method combining relational-logic representations with deep neural network learning. Domain-specific knowledge is described through relational rules which may be handcrafted or learned. The relational rule-set serves as a template for unfolding possibly deep neural networks whose structures also reflect the structure of given training or testing examples. Different networks corresponding to different examples share their weights, which co-evolve during training by stochastic gradient descend algorithm. Notable relational concepts can be discovered by interpreting shared hidden layer weights corresponding to the rules. Experiments on 78 relational learning benchmarks demonstrate the favorable performance of the method.