How would you approach improving the accuracy of a deep learning model that is underperforming?