A Unifying View on Implicit Bias in Training Linear Neural Networks