Dead pixel test using effective receptive field

Deep neural networks have been used in various fields, but their internal behavior is not well known. In this study, we discuss two counterintuitive behaviors of convolutional (CNNs). First, evaluated the size receptive field. Previous studies attempted to increase or control However, observed that field does describe classification accuracy. The would be inappropriate for representing superiority performance because it reflects only depth kernel and reflect other factors such as width cardinality. Second, using effective field, examined pixels contributing output. Intuitively, each pixel expected equally contribute final found there exist a partially dead state with little contribution We reveal reason lies architecture CNN solutions reduce phenomenon. Interestingly, general tasks, existence improves training CNNs. task captures small perturbation, degrade performance. Therefore, these should understood considered practical applications CNN.