Machine Unlearning with Minimal Gradient Dependence for High Unlearning Ratios