Fig. 1

The learnable weight matrix W can be decomposed into two matrices, Q and P. They are responsible for selecting target problem related tasks (AD vs NC, AD vs MCI, MCI vs NC) and features. By enforcing the l2−l1 norm of Q^T and P to be small, these group lasso penalty terms on rows on P and columns of Q encourage the rows of P and columns of Q in (1) to have all zero (rows and columns in gray) or non-zero elements. The first column of Q corresponds to the L-MCI vs E-MCI stratification task and the rest of them correspond to three auxiliary tasks. We could observe from the plot that the AD v NC and the AD vs MCI tasks are two related domains while the MCI vs NC task could not provide helpful information. Similarly, non-zero rows of P capture the shared features among useful domains