Distribution Benchmark

Distribution benchmark research focuses on evaluating the robustness of machine learning models to out-of-distribution (OOD) data—data differing significantly from training data. Current efforts concentrate on developing standardized benchmarks with diverse OOD scenarios (e.g., concept and covariate shifts, noise), exploring various OOD detection methods (including those leveraging test-time augmentation, nearest neighbor techniques, and prompt learning), and improving the accuracy and reliability of OOD detection scores. These advancements are crucial for building more reliable and safe AI systems across diverse applications, from image recognition and speech processing to drug discovery and medical diagnosis, where handling unexpected inputs is paramount.