Let's see now how Siamese network will learn to detect if two images are the same or not. First, we will fit as we solve both of the images:

And, in the end, we will gain the activation values in the last layer, and store those in memory of course.

Next, the derivation is calculated, and the feedback from that derivation will be back propagated to change the network weights in such a way that if the images are similar, then the difference of the encoded values will be small, while if the images are different, then the back propagation step will change the weights in such a way that will cause the difference of the activation values of the last layer or the encoded values to be large:

In a few words, this will be our goal of learning to ship the difference accordingly to small or large number depending if the images are the same or different.