Figure 2

Knowledge transfer in multi-task learning. Illustration of a knowledge transfer from task 2, which comprises a lot of training data (green), to a similar task 1, which contains little training data (blue). The ∊-insensitive tubes around the regression functions w₁^Tx and w₂^Tx are colored gray. The regularizer $J(w_1,\dots ,w_T)$ forces the model of task 1 (w₁) to be more similar to the model of task 2 (w₂). A model w₁ that is more similar to w₂ predicts the unknown data (red) better, which results in a better generalization of the model.