Hey wwc, thanks for the feedback. I’m glad you liked it !

You are right, the 4x4x2 vector doesn’t represent anything particular. It’s just a target vector that we are trying to reach starting from the 10x10x1 input vector. If that works then convolution on images should work too.

I would have gladly done a full example, but I would have needed two more types of layers : MaxPooling and Flatten.

If you think another article on how to code these two layers could be interesting, let me know !