Training Data Debugging Through a Lens of Outliers in Neuron Coverage

As the behavior of neural networks is dependent on the characteristics of training data, choosing appropriate data is mandatory to achieve expected levels of their prediction performance such as the accuracy or robustness. Data aug-mentation, or adding data, to improve accuracy can negatively affect robustness, because these two performance indicators are known to be incompatible. Training data debugging may involve deleting some data points as well as adding new ones; a question here is under what guiding principle the addition or deletion is conducted. Such a debugging method, debugging through a lens of outliers in neuron coverage, was proposed for cases of training classical fully-connected neural networks, but its applicability to CNN was unclear. This paper confirms, through experiments, that the debugging method is effective for cases of training CNN models as well, and furthermore demonstrates a concrete debugging process of discarding outliers and augmenting new data points so that the CNN exhibits balanced accuracy and robustness, without sacrificing neither performance indicator.