Simple, yet pretty effective - ~80% accuracy, when treating image as a flat array.
Simple network with 2 convolutional layers and 3 linear perceptrons. Trained using SGD on several (~15) epochs. Due to small batch size (4) it was very prone to sudden drop of accuracy, probably because large noise in sample data coming from small batch size.
Just an experiment, maybe a baseline for other models, it gets ~95%.
Just a different name, different layer parameters, nothing interested.
Potentially little bit better, more filters in conv layers