Extreme Classification Extreme Classification via Adversarial Softmax Approximation [ICLR 2020] Robert Bamler, Stephan Mandt. adversarial negative sample in extreme classification. auxiliary task to make negative sample is uniform by decision tree. proof the negative sample loss = constant + original loss. proof the best signal-to-nosie in the gradient is obtained is auxiliary model is best.