Python + Keras is a common method, but this time I made it from scratch with the familiar Visual Studio2022 and C # to understand the theory.

nnet_test1 [Visual Studio 2022, C#]

The simplest fully connected model. Enter the value of the upper left 4 squares. The black line learns and discriminates 4 patterns of horizontal, vertical, diagonal, or uniform.

1. The squares are grayscale (255 levels).
2. Convert the value (0-255) of each cell to float (0-0.9999) and input it to the input layer.
3. From the input layer to the middle layer. (green if the connection weight is negative, red if positive)
4. From the middle layer to the output layer. (horizon, vertical, cross, solid)
5. Update weights by error backpropagation.

There are 4 squares, but there are also 9 squares (3x3).

There are 15 middle layer, but there are also 50 or 100 petterns. There is also a 9 x 50 x 50 x 6 pattern with an additional middle layer.

Dynamically generated with supervised data including noise. 10,000 learnings. If you repeat it 10 times, you will get almost the correct answer.