Using 10 convolutional layers seems questionable - the precise values and location of each number is essential for getting the right answer, so it seems like th

Using 10 convolutional layers seems questionable - the precise values and location of each number is essential for getting the right answer, so it seems like the model design is poorly suited to the task. The obvious approach to me would be to train a neural net to be a heuristic to order exploration in a constraint satisfaction solver, mirroring how AlphaGo combines monte carlo search trees with neural nets. Might net a nice speed up.