ML Seminar: How exact can we make approximate Gaussian process inference?