Show HN: Adversarial Generation of Extreme Samples

It is interesting to mix EVT and GANs. Importance Splitting, a statistical tool that estimates probability of rare events and generates some of them, may also be closely related.

Great catch! Importance splitting is indeed a useful simulation technique for low probability events. The application of splitting based methods on bayesian models for extreme sample generation can be an interesting future work.

Extreme Value Theory is extensively used in anomaly detection as well. This can be a first step towards generating anomalous data which is usually quite difficult to find.

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.

Github Repository: https://github.com/Stream-AD/ExGAN