Linear

The neural networks underneath all our modern AI systems are in actuality piecewise linear approximators. When you train them on a custom loss function, you're making a polygon in Euclidean space that matches the shape of the loss' projection into the space. But still Euclidean.