Handwritten Digit Recognition Using Artificial Neural Networks (WIP)

Note: I am still writing the paper for this project, so this post is essentially just a repository of the results... for now.

Training set (10 sample images of each digit) from MNIST database

See: http://yann.lecun.com/exdb/mnist/

Training set transformed into a 100 x 784 image, 10 patterns per digit

Test set transformed into a 10000 x 784 image, 1000 patterns per digit.

Results

Training results

Convergence

Recognizer performance (heatmap)

Recognizer performance (full test set)