“According to research by Hestness et al. (2017), changes to a deep learning model's architecture affect the offset (E) of the power-law fit for generalization error, but do not change the exponent (α).”