Comparison of extracted rules from multiple networks

Neural networks can be trained to provide solutions in application domains where clear roles which would allow symbolic solutions do not exist. Neural networks in these domains still suffer from a major disadvantage, in that there is no explanation for why a particular decision was made by the network. The authors have generalised on their previous work on generating explanations for trained back-propagation neural networks to extract rules. The authors have found that there is significant variation in the quality of rules extracted from networks which have not been tuned for the task, and that the neural network correctness on the test set is not well correlated with the often better correctness of the extracted rules on the test set