Deep neural networks have become a popular tool for implementing intelligent systems. Their inherent black-box nature, however, remains a major drawback, compounded
by increasing network complexity which is expected to approach that of real brains in the near future. For safety verification and to build user and public confidence, it is crucial to comprehend how complex neural networks react to certain inputs, process information, and make decisions – especially in critical applications like autonomous driving or medicine. We develop diagnostic tools tailored to this purpose, partly inspired by research methods from cognitive neuroscience. As a demonstration, we show how a deep learning based
speech recognition system works internally.