sloppy
A basic neural network for recognizing handwritten numbers.
General
Sloppy, when trained with 60,000 and tested on 10.000 images, is able to recognize handwritten numbers by humans with an accuracy of ~96%.
This correctness may not be the same when using different training or testing data.
It is a simple, 3 layer neural network with 784 input and 10 output nodes. The number of hidden nodes may vary but use around 400-600 nodes for the best performance. Sigmoid is used as an activation function.
Sloppy can also recognize your handwriting but make sure you convert the image to 28x28 PNG and with a high contrast ratio between the background and foreground. This image will later be converted to a 1D array of 784 greyscale pixels.
A sample of training and testing data can be found in data.
If you want to train with more numbers download the training data with 60,000 sets and the testing data with 10,000 sets.
Contribute
If you have any questions or suggestions drop me with an email at mark@pearscom.com.