Possibilistic Predictive Uncertainty for Deep Learning (ICML 2026)

Yao Ni , Jeremie Houssineau , Yew-Soon Ong , Piotr Koniusz

Usage Guide

DAPPr.py provides a drop-in replacement for cross entropy and uncertainty scores for testing.

import torch.nn.functional as F
+from DAPPr import DAPPr_loss, DAPPr_uncertainty

# Training
for x, labels in train_loader:
    logits = model(x)
-   loss = F.cross_entropy(logits, labels)
+   loss = DAPPr_loss(logits, labels, lamb=2e-4)

# Testing: uncertainty estimation
for x, labels in test_loader:
    logits = model(x)
+   uncertainty = DAPPr_uncertainty(logits)
+   AU = uncertainty["AU"]  # aleatoric uncertainty
+   EU = uncertainty["EU"]  # epistemic uncertainty

lamb controls the regularization strength. In our experiments, we use either warm-up or linear scheduling for this value.

Environment

The code was tested with Python 3.11.7. Required Python packages are listed in requirements.txt.

Datasets

Please follow data/README.md to download and prepare the datasets.

Training for DAPPr

# CUB-200-2011
python3 train.py --dataset=CUB

# Stanford Dogs
python3 train.py --dataset=StanfordDogs 

# Tiny-ImageNet
python3 train.py --dataset=TinyImageNet --epochs=100 --lr=5e-3 --lamb=5e-3 --lamb_schedule=linear

Optional HDF5 Storage

The code supports both image-folder and HDF5 loading. Image-folder loading is the default.

If your filesystem is slow or has a file-count limit, follow HDF5 conversion instructions to convert image folders to HDF5. Then add --hdf5:

python3 train.py --dataset=CUB \
+ --hdf5

Running time

Reference running time on one V100 32GB GPU using image folders or HDF5 storage:

Dataset	Image Folders	HDF5
CUB-200-2011	1h50m	1h20m
StanfordDogs	2h40m	2h10m
TinyImageNet	10h40m	10h20m

Commands for EDL

python3 train.py --dataset=CUB --method=EDL --lamb=1e-5
python3 train.py --dataset=StanfordDogs --method=EDL --lamb=1e-5
python3 train.py --dataset=TinyImageNet --method=EDL --epochs=100 --lr=5e-3 --lamb=2e-4 --lamb_schedule=linear

Notes

Hyperparameter: For new datasets, we recommend tuning --lamb in the range 1e-5 to 1e-2. linear schedules usually require a larger value than warmup.

Entropy for OOD: For OOD detection, entropy is often stronger than epistemic uncertainty in our experiments. The code supports both.

Cite

@article{ni2026possibilistic,
  title={Possibilistic Predictive Uncertainty for Deep Learning},
  author={Ni, Yao and Houssineau, Jeremie and Ong, Yew Soon and Koniusz, Piotr},
  journal={arXiv preprint arXiv:2605.00600},
  year={2026}
}