ConvNet and PCA/SVM approaches to the kaggle digit recognizer challenge https://www.kaggle.com/c/digit-recognizer/data.
First approach uses a nonlinear SVM (ovo-shape) on the first 100 PCA components, done in Python with SciKitLearn (91% accuracy)
The second approach implements a Convolutional Neural Network with two hidden layers, done in Python with Tensorflow (99% accuracy)