The Effects of Using Results from Inversion by Evolutionary Algorithms to Retrain Artificial Neural Networks

The aim of inverting artificial neural networks (ANNs) is to find input patterns that are strongly classified as a predefined class. In this project an ANN is inverted by an evolutionary algorithm. The network is retrained by using the patterns extracted by the inversion as counter-examples, i.e. to classify the patterns as belonging to no class, which is the opposite of what the network previously did. The hypothesis is that the counter-examples extracted by the inversion will cause larger updates of the weights of the ANN and create a better mapping than what is caused by retraining using randomly generated counter-examples. This hypothesis is tested on recognition of pictures of handwritten digits. The tests indicate that this hypothesis is correct. However, the test- and training errors are higher when retraining using counter-examples, than for training only on examples of clean digits. It can be concluded that the counter-examples generated by the inversion have a great impact on the network. It is still unclear whether the quality of the network can be improved using this method