Extracting rules from trained feedforward neural networks with first order logic

Abstract

The need for neural-symbolic integration becomes evident as more complex problems are addressed, and they go beyond limited domain tasks such as classification. Search methods for extracting rules from neural networks work by sending input data combinations that activate a set of neurons. By properly ordering the input weights of a neuron, it is possible to narrow down the search space. Based on this observation, this work aims to present a method for extracting the pattern of rules learned by a trained feedforward neural network, analyzing its properties, and explaining these patterns through the use of first-order logic (FOL).