Do Neural Networks Recognize Patterns as well as Students?

We compare the similarity of humans and one of the current convolutional deep neural network (CNN) LeNet-5 on handwritten digits (MNIST database) recognition under stricter formulated recognition tasks and image degradations caused by a generative adversarial network (GAN). First, a conditional GAN neural network was trained on the MNIST database and generator seeds with normal probability distribution were used. The images for recognition were obtained by the generator that is part of this trained GAN network. Different image degradations were obtained by scaling the length of the input random vector for the generator. Second, as a criterion for the recognition task, we asked the question: Did a person write this digit (without thinning, thickening, protrusions or interruption of the line, no additional image artefacts which a human cannot write with a smooth stroke of the pen)? Human is more robust in most cases of recognizing degraded images. This is since a person who learns to recognize images throughout his life can better comply with the decision criterion than a neural network, which has a much simpler structure than the human brain and learns to recognize only on a limited training set. On the other hand, if we focus recognition criteria on the handwriting of digits, human is more precise, and CNN generalizes the recognised digit more behind this limitation. Our similarity graphs between CNNs and humans can be further used to find the boundaries of the handwritten digits domain and to create a system for recognizing anomalous images.