Abstract
In recent years, deep neural networks have revolutionized many applicationdomains of machine learning and are key components of many critical decision orpredictive processes. Therefore, it is crucial that domain specialists canunderstand and analyze actions and pre- dictions, even of the most complexneural network architectures. Despite these arguments neural networks are oftentreated as black boxes. In the attempt to alleviate this short- coming manyanalysis methods were proposed, yet the lack of reference implementations oftenmakes a systematic comparison between the methods a major effort. The presentedlibrary iNNvestigate addresses this by providing a common interface andout-of-the- box implementation for many analysis methods, including thereference implementation for PatternNet and PatternAttribution as well as forLRP-methods. To demonstrate the versatility of iNNvestigate, we provide ananalysis of image classifications for variety of state-of-the-art neuralnetwork architectures.