RAD Framework Detects Ambiguity in Machine Learning Predictions

Machine learning models are expected to be robust, meaning their predictions should remain consistent even with minor changes to the model or its inputs. When predictions change significantly under such permissible variations, they are considered "ambiguous," indicating a lack of reliability. Identifying this ambiguity in deployed models is challenging but vital, particularly in critical applications. The new Robust Ambiguity Detection (RAD) framework offers a solution by quantifying predictive ambiguity through two distinct metrics: Model-Space Consistency and Feature-Space Consistency. These metrics form a "RAD Score-Pair" which, when visualized in a RAD Plot, provides an interpretable understanding of the sources of ambiguity. The framework enables models to abstain from ambiguous predictions or flag them for human intervention, enhancing trustworthiness. Evaluated on both synthetic and real-world datasets, RAD demonstrates its ability to identify ambiguity effectively, with a downstream application showing comparable performance to existing rejection-based methods for abstaining from the most ambiguous predictions.