A comparison between one-vs-all logistic regression classification and NeuralNetwork with backpropagation method to train a model for MNIST dataset. With the Backpropgation an accuracy of >96% is achieved, whicle the one-vs-all method gives >80% accuracy.
See NeuralNet.ipynb for different methods on smaller set of data and full-mnist.ipynb for the regularized backpropagation method on the full MNIST dataset.