Asymmetric Transfer

A transfer learning phenomenon in machine learning in which different types of training data exhibit unequal generalization effects across downstream tasks. One category of data broadly improves performance on a wide range of tasks, while another category's transferability is comparatively limited.