A Step towards Explainable Artificial NeuralNetworks in Image Processingby Dataset Assessment

We propose a methodology for generalized exploratory
data analysis focusing on artificial neural network
(ANN) methods. Our method is denoted IC-ACC due to the
combined assessment of information content (IC) and accuracy
(ACC) and aims at answering a frequently posed question
in ANN research: ”What is good data?” As the dataset
has the primary influence on the development of the model,
IC-ACC provides a step towards explainable ANN methods
in the pre-modeling stage by a better insight in the dataset.
With this insight, detrimental data can be eliminated before a
negative influence on the ANN performance occurs. IC-ACC
constitutes a guideline to generate efficient and accurate data
for a specific, data-driven ANN method. Moreover, we show
that training an ANN for the semantic segmentation of 3D
data from unstructured environments with IC-ACC-assessed
and -customized training data contributes to a more efficient
training. The IC-ACC method is demonstrated on application
examples for the visual perception of robotic platforms.