Skip to main content
Back

DAPPr Fixes AI Overconfidence With a Forgotten 1970s Math Theory

DAPPr uses possibility theory to give neural networks principled, cheap uncertainty estimates — beating Bayesian methods and EDL variants across benchmarks

  • DAPPr introduces possibility theory to deep learning uncertainty, replacing expensive Bayesian integration with tractable supremum operations.
  • The method is a drop-in cross-entropy replacement (~10 lines of PyTorch) requiring no architectural changes and one small hyperparameter.
  • Unlike EDL, DAPPr's epistemic uncertainty provably decreases with more data, fixing a known pathological failure mode of evidential methods.
  • A single DAPPr model outperforms 10-model ensembles on fine-grained datasets (CUB-200, Stanford Dogs) for accuracy and OOD detection.
  • Evaluated across MNIST, CIFAR-10/100, long-tailed, distribution-shift, and fine-grained benchmarks — best or second-best on nearly all metrics.
  • Code is open-source and the paper is accepted at ICML.

Deep neural networks have a well-known overconfidence problem: ask a model about something it has never seen, and it will often answer with high certainty anyway. Fixing this requires epistemic uncertainty quantification , the ability to say "I don't know" when the model genuinely lacks knowledge. A new paper accepted at ICML introduces DAPPr (Dirichlet-approximated possibilistic posterior predictions), a framework that resolves a long-standing tension in this field by borrowing from a largely forgotten mathematical theory.

The dilemma nobody had solved

Two camps have dominated epistemic uncertainty research, and both have serious drawbacks. Bayesian approaches provide principled estimates but remain computationally prohibitive, while efficient second-order predictors lack rigorous derivations connecting their specific objectives to epistemic uncertainty quantification.

To unpack that: Bayesian methods like MC Dropout or deep ensembles work by training multiple models and measuring how much they disagree. That disagreement is a principled signal for uncertainty, but it requires running inference through many networks simultaneously , expensive at scale. The cheaper alternative, second-order predictors (methods that output a distribution over predictions rather than a single prediction), sidestep this cost by having one network output the parameters of a Dirichlet distribution. The most prominent of these is Evidential Deep Learning (EDL). But EDL and its variants have a dirty secret: recent analyses highlight identifiability and convergence pathologies, showing that the resulting epistemic uncertainty estimates are tightly controlled by regularization, making them only meaningful in a relative sense. In fact, EDL can exhibit a pathological behavior where its reported uncertainty increases as you feed it more training data , the opposite of what should happen.

Possibility theory: the third path

The authors turn to possibility theory, a mathematical framework introduced in the 1970s that has remained largely unexplored in deep learning. Unlike probability theory, which requires all probabilities to sum to one (forcing a complete allocation of belief), possibility theory uses