On Penalty-based Bilevel Gradient Descent Method