Code for the Digit Recognizer competition at Kaggle
In this folder, kernel approximation code from Paul Mineiro. In
maxout, maxout code from Pylearn2 for 98.3% accuracy.
Paul Mineiro's cosplay
Paul Mineiro published some kernel approximation code, dubbed cosplay for use with MNIST. Here's both the original and a version modified to output predictions for the Kaggle Digit Recognizer competition, which is also based on MNIST.
As mentioned in the article, you'll need Maxent and L-BFGS for the logistic regression part.
Accuracy on Digits is similar to accuracy on MNIST: around 97.7%.