RMNIST/N: Train MNIST dataset with only TEN samples

RMNIST/N is a dataset that reduces MNIST with N examples for each digit class. In this way, RMNIST/1 has 1 training example for each digit, for a total of only 10 training examples. This article presents how a digits recognizer can be trained on only ten samples from the whole MNIST dataset. Data augmentation is used during the training. Ablation study is also made.