Input complexity and out-of-distribution detection with likelihood-based generative models

Likelihood-based generative models are a promising resource to detect\nout-of-distribution (OOD) inputs which could compromise the robustness or\nreliability of a machine learning system. However, likelihoods derived from\nsuch models have been shown to be problematic for detecting certain types of\ninputs that significantly differ from training data. In this paper, we pose\nthat this problem is due to the excessive influence that input complexity has\nin generative models' likelihoods. We report a set of experiments supporting\nthis hypothesis, and use an estimate of input complexity to derive an efficient\nand parameter-free OOD score, which can be seen as a likelihood-ratio, akin to\nBayesian model comparison. We find such score to perform comparably to, or even\nbetter than, existing OOD detection approaches under a wide range of data sets,\nmodels, model sizes, and complexity estimates.\n