Deep Relational Learning

keywords: Statistical Relational Learning, Relational Logic, Templated Models, Lifted Modeling, Deep Learning, Neural Networks

abstract: In this proposal, we strive to marry rich, interpretable, logic-based representations of statistical relational learning with efficient, deep neural network training, under the paradigm of lifted modeling approaches, to leverage the praised advantages of both symbolic and statistical machine learning worlds. We believe that this strategy will rendered us able to create deep learning architectures capable of learning from graphs and otherwise complex relational data, incorporating background expert knowledge in the expressive form of logical theories, and interpret the learning results of neural networks. Preliminary experiments on structured learning benchmarks suggest for feasibility of the proposed strategy.