Subjective Risk Decomposition: A New View for Uncertainty Quantification

arXiv cs.AI · 6 hours ago

Researchers propose a framework where uncertainty measures in machine learning emerge from decomposing subjective risk based on strictly proper loss functions rather than being treated as fundamental primitives. The approach recovers existing epistemic and aleatoric uncertainty measures including reverse cross-entropy through information-theoretic decomposition. This theoretical foundation suggests practitioners can derive appropriate uncertainty quantification terms directly from their chosen loss function and modeling scenario.