This neural network takes in a $28\times28$ sized grid. It is stored within as $1$s and $0$s, corresponding to that grid's light-levels. Then, it is pooled into 2 dimensions, through 2-dimensional max pooling, from $28\times28$ to $14\times14$. This was done as the original context of the problem was found on DMOJ, the judge does not accept anything more than $63000$ characters. I trained it myself, and on similar to MNIST data, it recieves an average accuracy of $95.5\%$. Perhaps it is due to this demonstrations nature, however, such accuracy is not very possible, unless you are to draw an anatomically perfect number. It could be completely wrong too, as during its learning, it was trained to prioritize confidence, with a "rather be wrong and confident than right and unconfident".