Existing GAN based approaches excel at generating realistic samples, but seek to generate typical samples, rather than extreme samples. We propose ExGAN to generate realistic and extreme samples.
ExGAN allows the user to specify both the desired extremeness measure, as well as the desired extremeness probability they wish to sample at. Generating increasingly extreme examples can be done in constant time (with respect to the extremeness probability), as opposed to the exponential time required by the baseline.
ExGAN looks cool! My initial thinking was incorporating it into texture synthesis. Most neural based image generators can't sample intelligently beyond the input example. But the "extremeness measure" appears to be much more robust. And can be used in simulating any world model beyond historic regimes ;)
ExGAN allows the user to specify both the desired extremeness measure, as well as the desired extremeness probability they wish to sample at. Generating increasingly extreme examples can be done in constant time (with respect to the extremeness probability), as opposed to the exponential time required by the baseline.
Paper: https://arxiv.org/abs/2009.08454