High-dimensional Asymptotics of Feature Learning: How One Gradient Step Improves the Representation Jimmy Ba1, Murat A. Erdogdu 1, Taiji Suzuki