1_dLYdRMQ96L_Uhi-BUBPsig

adversary is just trying to minimize its own loss, we update its weights U according to the gradient ∇_uL_A as usual. However, the update expression for the predictor’s weights W gets a little more complicated