Reconstructing Deep Neural Networks: Unleashing the Optimization Potential of Natural Gradient Descent