Unsupervised Minimax: Adversarial Curiosity, Generative Adversarial Networks, and Predictability Minimization A new review paper by Schmidhuber linking GANs to earlier ideas about unsupervised minimax games. Might be useful for those looking beyond GANs. [arxiv.org][twitter]PM uses gradient-based minimax or adversarial training to learn an encoder of the data, such that the codes are distributed like the data, and the probability of a given pattern can be read off its code as the product of the predictor-modeled probabilities of the code components (Sec