Interpretation of Neural Networks for Classification Tasks

To overcome the black box behaviour of neural networks many different approaches have been proposed. Up to now there is no standard tool being able to handle general feedforward networks. In this paper a method is proposed to combine visualization techniques with transformation algorithms to interpret neural feedforward networks. An application in ultrasonic crack detection will show that the overcoming of the black box structure of neural networks is not academical. It shows that the optimization of a network after training is possible and highly usable.