Shifts: A Dataset of Real Distributional Shift Across Multiple Large-Scale Tasks

There has been significant research done on developing methods for improving\nrobustness to distributional shift and uncertainty estimation. In contrast,\nonly limited work has examined developing standard datasets and benchmarks for\nassessing these approaches. Additionally, most work on uncertainty estimation\nand robustness has developed new techniques based on small-scale regression or\nimage classification tasks. However, many tasks of practical interest have\ndifferent modalities, such as tabular data, audio, text, or sensor data, which\noffer significant challenges involving regression and discrete or continuous\nstructured prediction. Thus, given the current state of the field, a\nstandardized large-scale dataset of tasks across a range of modalities affected\nby distributional shifts is necessary. This will enable researchers to\nmeaningfully evaluate the plethora of recently developed uncertainty\nquantification methods, as well as assessment criteria and state-of-the-art\nbaselines. In this work, we propose the Shifts Dataset for evaluation of\nuncertainty estimates and robustness to distributional shift. The dataset,\nwhich has been collected from industrial sources and services, is composed of\nthree tasks, with each corresponding to a particular data modality: tabular\nweather prediction, machine translation, and self-driving car (SDC) vehicle\nmotion prediction. All of these data modalities and tasks are affected by real,\n"in-the-wild" distributional shifts and pose interesting challenges with\nrespect to uncertainty estimation. In this work we provide a description of the\ndataset and baseline results for all tasks.\n