In this project we tried to compare the performance of 3 different methods for handwritten digit recognition using the famous MNIST DataSet.
I have tried to use Neural Networks, SVM and SVM plus Principal component Analysis.
A SVM model with kernel of 9th order, gamma = 0.01 and cost = 1000 gave me a test set accuracy of 95.24%.
A Neural net model with rectifier activation gave a test set accuracy of 97.41%. This might be further improved.My hyper parameter tuning was limited by my hardware.
A SVM model on a PCA transformed data gave the best result of 98%.This is inutitive as PCA would transform the dataset into more relevant features and the less important features could be dropped. I also performed a grid search on the data for the most relevant combination of hyper parameters.
Please run the code and play along for further insight.