This project is an attempt to beat the benchmarks set for the algorithms described in the 1998 paper by LeCun et al., plus some newer algorithms that have become popular, using the MNIST database of handwritten digits. I then tried each algorithm in the Kaggle Handwritten Digit contest.
As of August 2015 I have gotten through all of the algorithms and achieved 0.99257 on Kaggle using a Theano multi-layer neural network running on a GPU at AWS, placing 28th out of 646 entrants.
See MNIST.pdf for the detailed documentation.
