MNIST trainer written in PyTorch. Please see the notebook mnist.ipynb for details, or run it yourself.
Our statistically thorough result claims 99.82 percent accuracy on the test set, but if we throw in all the nets we trained into one big committee, we get 99.84 percent accuracy. On the ten thousand test images, these sixteen are the ones that are misclassified.
