OpenGAN: Open-Set Recognition via Open Data Generation

D fake features K-way softmax closed vs. fake closed-set images G D G closed-set images outlier images D outlier images closed-set images D closed vs. open fake images outlier images closed-set images fake images closed vs. open G closed vs. open (a) GAN (b) Outlier Exposure (c) OpenGAN pix (d) OpenGAN f ea Figure 1: We explore open-set recognition, which requires the ability to discriminate open-set test examples outside K classes of interest. (a) Past work has suggested that GAN discriminators can serve as open-set likelihood functions, but this does not work well due to instable training of GANs [56, 53, 48, 66, 37]. (b) Outlier Exposure [31] exploits some outlier data to learn a binary discriminator D for openset discrimination. Because outliers observed during training will not exhaustively span the open-world, the discriminator D tends to generalize poorly to diverse open data [57]. (c) We introduce OpenGAN, which augments training outliers with fake open data synthesized by a generator G trained to fool the discriminator D. Importantly, we find that a small number of outliers stabilizes training by enabling effective model selection of the discriminator D. (d) Because we are concerned with accurate discrimination rather than realistic pixel generation, we find it more efficient to generate (and discriminate) features from the off-the-shelf K-way classification network. This allows OpenGAN to be implemented via a lightweight discriminator head built on top of an existing K-way network, enabling closedworld systems to be readily modified for open-set recognition.