SIPTA Probly Challenge
CORRUPTED MNIST · UNCERTAINTY DECOMPOSITION

Identify corrupted
observations.

Analyze the provided set of … MNIST images and identify the hidden corruptions. Use probly to decompose predictive uncertainty and compare epistemic, aleatoric, and total uncertainty signals.

Analysis libraryproblySCAN TO OPENsipta2026.cortys.deQR code for sipta2026.cortys.de
CHALLENGE ACCESS

Start or continue your work

No account is required. A private group access link is generated after registration. New groups receive the configured number of submission attempts.

LeaderboardView current ranksGoogle ColabOpen the notebookGitHub repositoryUse a local environment
—image IDs
—hidden outliers
—attempts per group

How the challenge works

Open the notebook, develop an approach for detecting outliers in the provided dataset, and submit the IDs of the observations you identify as outliers.

01
</>

Open the notebook

Start from the shared dataset and environment in Google Colab or the repository.

02
≈≠≈

Detect the outliers

Develop and train an approach for identifying the corrupted observations, ideally using probly.

03
→ ID

Submit the IDs

Submit the predicted outlier IDs from your group workspace and review the resulting rank.

Group leaderboard

Full leaderboard
◇No groups registered

Registered groups will appear here.