Changing the Training Data Distribution to Reduce Simplicity Bias Improves In-distribution Generalization

Open in new window