Towards a Statistical Understanding of Neural Networks: Beyond the Neural Tangent Kernel Theories