Relational knowledge extraction from neural networks
Manoel Vitor Macedo França, Artur d’Avila Garcez, Gerson Zaverucha · Neural Information Processing Systems · 2015
The effective integration of learning and reasoning is a well-known and challenging area of research within artificial intelligence. Neural-symbolic systems seek to integrate learning and reasoning by combining neural networks and symbolic knowledge representation. In this paper, a novel methodology is proposed for the extraction of relational knowledge from neural networks which are trainable by the efficient application of the backpropagation learning algorithm. First-order logic rules are extracted from the neural networks, offering interpretable symbolic relational models on which logical reasoning can be performed. The wellknown knowledge extraction algorithm TREPAN was adapted and incorporated into the first-order version of the neural-symbolic system CILP++. Empirical results obtained in comparison with a probabilistic model for relational learning, Markov Logic Networks, and a state-of-the-art Inductive Logic Programming system, Aleph, indicate that the proposed methodology achieves competitive accuracy results consistently in all datasets investigated, while either Markov Logic Networks or Aleph show considerably worse results in at least one dataset. It is expected that effective knowledge extraction from neural networks can contribute to the integration of heterogeneous knowledge representations.