A Classification model trained entirely on a scientific calculator [P]

The calculator model is the Casio FX-82CE X. It is not programmable or graphical so everything had to be done by hand.

The architecture is simple, MNIST images (just 0s and 1s) downscaled to 3x3, with binary pixels, then with a fully connected layer, they are brought down to just 1 neuron, which serves as the output neuron. If it's value is above 0 it counts as a one, otherwise as a zero.

The training was done with a simple perceptron style training without a bias, 6 images (3 per class) and these are the final weights; "0 0 1 -1 2 -1 -1 1 -1"

I tested it's accuracy on my phone, using the validation segment of mnist, and it got 67.04% validation accuracy on this binary classification task. Interestingly it predicted every "one" correctly but predicted most of the "zeros" incorrectly.

Finally I decided to see what this architecture's max could potentially be with sufficient training: after 1000 epochs on the 0s and 1s in the training split it got a validation accuracy of 98.96%, the training also slightly differs from the manual one as this one uses SGD instead of the perceptron style approach. These are the final weights for this run; -0.509 -4.451 0.086 -5.775 10.651 -7.630 -0.012 -2.560 2.304