Task-Agnostic Out-of-Distribution Detection Using Kernel Density Estimation

In the recent years, researchers proposed a number of successful methods to perform out-of-distribution (OOD) detection in deep neural networks (DNNs). So far scope highly accurate has been limited image level classification tasks. However, attempts for generally applicable beyond did not attain similar performance. this paper, we address limitation by proposing simple yet effective task-agnostic OOD method. We estimate probability density functions (pdfs) intermediate features pre-trained DNN performing kernel estimation (KDE) on training dataset. As direct application KDE feature maps is hindered their high dimensionality, use set lower-dimensional marginalized models instead single high-dimensional one. At test time, evaluate pdfs sample and produce confidence score that indicates OOD. The eliminates need making simplifying assumptions about underlying makes method task-agnostic. experiments task using computer vision benchmark datasets. Additionally, medical segmentation brain MRI results demonstrate consistently achieves performance both tasks improves state-of-the-art almost all cases. Our code available at https://github.com/eerdil/task_agnostic_ood. Longer version paper supplementary materials can be found as preprint [8].