Notes. We compare the results with and without transfer learning on m12f LSR0. When the transfer learning is performed, either the first layer is updated, or the last layer is updated, or both. In the transfer learning step, all truth labels are derived using either the VM or ZM selections, as defined in Sect. 3. ∑χ² quantifies how similar the v_R, v_ϕ, and v_z are to the truth-level distributions, as illustrated in Fig. 8. For each transfer learning method, we determine the optimal cut on the network score, as described in the text. In general, the ZM selection with transfer learning performed by only varying the last layer does the best job at reproducing the kinematic distributions of the true accreted stars, that is, it achieves the lowest ∑χ².