How do I improve my random forest prediction rate from 96%




While participating in the Kaggle competition on Hand written digit recognition I could achieve an accuracy rate of 96%(50th percentile) using a combination of the train - test data and generating 3 random forests and combining them for prediction.
But I want to improve more on this model.I had tried rotating the images by 5-10 degrees but R gives memory usage error.
Is there any other way to improve the model performance.


Have you tried deep neural nets? They might be of help here.

Also, if your laptop is giving memory errors, you can always spin up a machine on AWS :smile: