Preliminary results: up to 90% on hand drawn image classification (784 pixels, 28x28). Thousands of images used for training. Tested on thousands of images also.

Up to 91.59% on 11 sanskrit alphabets (handwritten 32x32 = 1024 px images. NeuralNet topology 1 input, 2 hid, 1 output).

ETC..