Speaker
Description
The utilization of artificial intelligence in critical and high-risk applications is poised to become a prevalent practice soon and the development of a methodology for testing and approving machine or deep learning models is imperative. As most experience has been accumulated in the medical field to date, this contribution employs a classification task using electroencephalography data to illustrate the preparation of reliable and effective approval and test methods.A deep learning network was trained to assign auditory brainstem responses to three loudness classes. The generation of artificial data was achieved by utilizing audiological experimental reference data, with a local instance being defined to represent the normal hearing of a hypothetical test subject. The time series (trials) were created by employing a perturbation method. The single-layer perceptron model was selected amenable to an interpretation, with the weights indicating the data to which the network is sensitive. To validate a specified relevance parameter derived from the perceptron weights, the accuracy and a metric distance were calculated as performance measures. It has been demonstrated that the perceptron can characterize the relevance of specific data within a local environment surrounding the instance. Furthermore, conclusions have been drawn regarding the rendering of new test data.