Compact visualization of DNN classification performances for interpretation and improvement

In the research field of automatic classification tasks, deep neural networks (or DNN) are frequently used for their efficiency and adaptive nature. State-of-the-art architectures and pre-trained networks exist and can be converted and fine tuned to handle other classification tasks. However, because of their training phase requiring many computations and adjustments on the many parameters of the models, they suffer from a black-box effect that