This is the second coding part in the whole project, which is conducted in R.
We introduce a novel OOD detection method based on Gaussian processes, focusing on establishing a detection boundary using only in-distribution (InD) samples. The core concept revolves around quantifying uncertainty in unconstrained softmax scores through a clustered Gaussian process. By leveraging the posterior predictive distribution from the Gaussian process, we define a score function that effectively distinguishes between InD and OOD samples.
I am mentored by Dr. Wenbo Sun, Dr. Arpan Kusari from University of Michgan and Prof. Chih-Li Sung from Michigan State University.
Download the repo and run the KL.Rmd file.