March 22, 2019
At the special exhibition "Smart New World" at the Science Center phaeno in Wolfsburg, the Fraunhofer Heinrich Hertz Institute presents LRP demos of age and gender determination using facial recognition as well as number recognition and traceability using "LRP" (Layerwise Relevance Propagation).
The special exhibition "Smart New World" starts on March 22. The exhibition consists of several stations at which knowledge on the topics of coding, Artificial Intelligence, Internet of Things and social networks is conveyed in a playful manner. The participatory stations invite us to reflect on how digitization affects our lives and ourselves.
Fraunhofer HHI presents its LRP demo for explanation. Through this Fraunhofer HHI program, facial classification, handwriting, image and text recognition as well as Visual Question Answering can be reconstructed. Thus, the computer program can determine age and gender using an artificial neural network. In order to perform this function, the artificial neural network was trained with a large number of images of faces. LRP provides an effective method for verifying this process. This involves a step-by-step tracing of the image areas relevant for the decision. This allows an assessment whether there is a plausible relationship between the determined image areas and the recognition.
Towards the other function of number recognition the neural nets were trained only with numbers. In this case, a heat map provides a means of reproducing recognition in artificial neural networks. This method can be applied to track how, for example, a "3" is slowly converted into an "8". This way, deeper insights into the operation methods of artificial neural networks could already be offered. With this it was possible to demonstrate, for example, that an Artificial Intelligence did not recognize horses on the basis of the horses in the pictures, but on the basis of the copyright information in the picture.