Appendix to " Auxiliary Task Reweighting for Minimum-data Learning " Anonymous Author(s) Affiliation Address email 1 Additional Discussion on ARML 1

Appendix to "Auxiliary T ask Reweighting for Minimum-data Learning" S is the optimal parameter. Concretely, with Jensen's inequality, we have Z ( S; α) is small). See the remark below for more discussion on the validity of the assumption. In Section 2.4, we propose a two-stage algorithm, where we update the task weights with Langevin This simplified version is summarized in Algorithm 1. For all results, we repeat experiments for three times and report the average performance.