Draw a SINGLE digit (0-9) in the box at a reasonable size (not tiny), click Enter, and this neural network will output what digit it thinks you most likely drew. The smaller box represents what the neural network sees as input (a lower resolution of your drawing which has been centred and resized to fill the whole box) Currently this is far from perfect. The network seems to struggle with identifying some numbers more than others. I will continue to train the neural network further to increase accuracy as I expand the training data. Let me know which numbers it struggles on and I can refine the training data based on those weaknesses when I next train the network. For those interested, the network design is a Multilayer Perceptron that takes a 16x16 pixel representation of the number as input which is fed through a 128 neuron hidden layer, and then to the output layer (10 neurons for the 10 possible digits). It uses ReLU for the hidden activation function and Softmax for the output. It is trained through backpropagation. I train the network on a separate scratch project and then import the weights and biases into this one to update the network.
