Dirichlet-Approximated Possibilistic Posterior Predictions for Deep Learning Uncertainty
A new framework called Dirichlet-approximated possibilistic posterior predictions (DAPPr) addresses the dilemma in epistemic uncertainty modelling for deep neural networks. Bayesian approaches offer principled estimates but are computationally prohibitive, while efficient second-order predictors lack rigorous derivations. DAPPr leverages possibility theory to define a possibilistic posterior over parameters, projects it to prediction space via supremum operators, and approximates it with learnable Dirichlet possibility functions. This yields a simple training objective with closed-form solutions, enabling reliable uncertainty quantification without heavy computation. The method is introduced in arXiv:2605.00600.
Key facts
- DAPPr stands for Dirichlet-approximated possibilistic posterior predictions.
- It is a framework for epistemic uncertainty modelling in deep neural networks.
- Bayesian approaches are principled but computationally prohibitive.
- Efficient second-order predictors lack rigorous derivations.
- DAPPr uses possibility theory to define a possibilistic posterior over parameters.
- The posterior is projected to prediction space via supremum operators.
- The projected posterior is approximated using learnable Dirichlet possibility functions.
- The training objective has closed-form solutions.
- The method is introduced in arXiv:2605.00600.
Entities
Institutions
- arXiv