This CNN-based model uses data augmentation and dropout to get a highly accurate handwriting digit recognition on the MNIST dataset. The model uses Tensorflow and Keras functionality (data augmentation, callbacks, etc.) and is GPU-enabled.
What differentiates this model from many other MNIST models is it's use of data augmentation (small rotations and shifts) and early stopping to control overfitting and improve generalization. The approach is also unique in that it does not use batch normalization to address overfitting, instead relying solely on dropout to do so. The final unique aspect is to keep the number of parameters as small as possible, while still achieving high accuracy. This goal of having a small neural net (i.e. fewer parameters) is important if predictive models of computer vision are used in edge computing.