We investigate the behavior of convolutional neural networks (CNN) in the presence of label noise. We show empirically that CNN prediction for a given test sample depends on the labels of the training samples in its local neighborhood. This is similar to the way that the →