On the Traceability of Results from Deep Learning-based Cloud Services

Deep learning-based approaches have become an important method for media content analysis, and are useful tools for multimedia analytics, as they enable organising and visualising multimedia content items. However, the use of deep neural networks also raises issues of traceability, reproducability and understanding analysis results. The issues are caused by the dependency on training data sets and their possible bias, the change of training data sets over time and the lack of transparent and interoperable representations of models. In this paper we analyse these problems in detail and provide examples. We propose six recommendations to address these issues, which include having interoperable representations of trained models, the identification of training data and models (including versions) and the description of provenance of data sets, models and results.