Epistemic uncertainty quantification is a crucial part of drawing credible conclusions from predictive models, whether concerned about the prediction at a given point or any downstream evaluation that uses the model as input.

Specifically, we first incorporate the instance-wise alignment module by maximizing the agreement between image-report pairs.

This is, e.g., the typical setting of many ERM

Each dataset has 513 instances on average. School contains the examination scores of students from 139 schools (datasets). The average outlier rates in a dataset of IoT, Landmine, and School are 0.05, 0.06, and 0.15, For IoT, we randomly chose one target, one validation, and seven source datasets. Algorithm 1 Training procedure of our model for inlier-based outlier detection.Require: LOF and IF use only target unlabeled instances to find outliers. For both methods, the best test AUCs were reported.

This is enabled through a new theoretical technique to analyze the Rényi Differential Privacy (RDP) of the sub-sampled shuffle model.