2 ms·

So it doesn't use a neural network, but it is still optimized by gradient descent. Differentiability is the key!