Out of distribution detection with DLSGAN

DLSGAN proposed a learning-based GAN inversion method with maximum likelihood estimation. In this paper, I propose a method for out-of-distribution detection using the encoder of DLSGAN. Simply, the log-likelihood of the predicted latent code of input data can be used for out-of-distribution (OOD) detection. 1. OOD detection DLSGAN DLSGAN [4] proposed a learning-based GAN inversion method with maximum likelihood estimation of the encoder. The encoder of DLSGAN maps input data to predicted latent code. When the DLSGAN converged, one can know the true distribution of DLSGAN encoder output. Therefore, the log-likelihood of input data can be simply calculated through the DLSGAN encoder. The following equation shows the log-likelihood of the predicted latent code of input data. ood score = sum(log f(E(x)|μ, v)) In the above equation, x and E represent input data and DLSGAN encoder, respectively. E(x) represents d_z -dimensional predicted latent code of input data x . f represents probability density function of the i.i.d. latent random variable Z. μ and v represents mean and variance vector for the probability density function f. μ is mean vector of latent random variable Z . v is the same vector as traced variance vector of DLSGAN. The ood score is simply the log-likelihood of the predicted latent code E(x). If the ood score is smaller than the threshold, the input data is classified as OOD data. Otherwise, it is classified as in-distribution data.