No. There are no neural networks here. The renderer is just a function that takes a bunch of ellipsoid parameters and outputs a bunch of pixels. You render the scene, then subtract the ground truth pixels from the result, and sum the squared differences to get the total error. Then you ask the question "how would the error change if the X position of ellipsoid #1 was changed slightly?" (then repeat for all ellipsoid parameters, not just the X position, and all ellipsoids, not just ellipsoid #1). In other words, compute the partial derivative of the error with respect to each ellipsoid parameter. This gives you a gradient, that you can use to adjust the ellipsoids to decrease the error (i.e. get closer to the ground truth image).