RGB Images

This experiment uses the default configuration with 128 neurons in the dense layer. The input images were scaled down to 32x32 pixels but all three colour channels remain intact.

After 1000 epochs of training it recognises all training images and ~90% of the test images.

Here are some of the mislabelled images. Note how it confuses numbers with similar colours, for instance 4 and 9. This suggests that the network figured out connection between colour and number, but not the connection between shape and number.