Epistemic Uncertainty in Conformal Scores: A Unified Approach Conformal prediction methods create prediction bands with distribution-free guarantees but do not explicitly capture epistemic uncertainty, which can lead to overconfident predictions in data-sparse r...

5/
EPICSCORE is accurate, flexible, and broadly applicable 💥
📄 Paper: arxiv.org/abs/2502.06995
#AI #ML #UncertaintyQuantification #ConformalPrediction #BayesianMethods

4/
Strong results across tasks 📈
EPICSCORE adapts well to diverse settings—from regression to image classification—while improving uncertainty estimates 🔍✅

3/
We use Bayesian models—BART, GPs, MC Dropout— to adapt the interval width depending on data availability.

📊 More data → tighter intervals
🌌 Less data → wider intervals (epistemic uncertainty!)

2/
We built EPICSCORE:
✅ Works with any conformal score
✅ Adds Bayesian modeling of epistemic uncertainty
✅ Keeps all coverage guarantees
✅ Achieves asymptotic conditional coverage

1/
Conformal prediction is great for distribution-free coverage ✅
But it misses epistemic uncertainty—when data is sparse, and the model just doesn’t know 🤷

Most fixes are task-specific (regression, quantile) and hard to generalize.

Big news—our paper was accepted for oral presentation at #UAI2025! 🥳

We (me + Vagner S. Santos, Thiago R Ramos, @rafael-izbicki.bsky.social ) built EPICSCORE, a method to improve conformal prediction intervals by adding epistemic uncertainty awareness.

Here’s why it matters 🧵