Eliminating the Invariance on the Loss Landscape of Linear Autoencoders