HAICON 2026 Workshop on Neural Network Uncertainty Quantification: Making Model Predictions More Trustworthy

Official workshop materials for HAICON 2026, systematically explaining uncertainty quantification methods in deep learning from theory to practice, covering regression, classification, and Earth observation applications, and providing directly runnable Colab notebooks.

• Original Author/Maintainer: ChrisKo94
• Source Platform: GitHub
• Original Title: HAICON_2026_Trustworthy_UQ_Workshop
• Original Link: https://github.com/ChrisKo94/HAICON_2026_Trustworthy_UQ_Workshop
• Publication Date: 2026-06-04

Deep learning models have achieved remarkable results in fields like image recognition, natural language processing, and autonomous driving, but a fundamental question has always plagued practitioners: Does the model know what it doesn't know?

Imagine a medical AI system diagnosing a patient's condition, or an autonomous vehicle identifying road obstacles. If the model is overly confident in wrong predictions, the consequences could be catastrophic. This is the core value of Uncertainty Quantification (UQ) — enabling models to honestly say "I'm not sure".

HAICON 2026 (AI for Health, Environment and Earth Sciences Conference) specifically organized a one-hour hands-on workshop for this purpose, covering UQ theory and practice comprehensively from toy regression examples to real-world Earth observation applications.

The entire workshop consists of four progressive Jupyter Notebooks, each with a directly runnable Google Colab link:

This section starts with the simplest regression problems and introduces multiple basic yet effective methods:

• Gaussian Negative Log-Likelihood (Gaussian NLL): Not just predicting a single value, but predicting both mean and variance simultaneously, allowing the model to learn to express data noise
• Model Calibration: Checking whether the model's predicted confidence matches the actual accuracy, addressing the "overconfidence" issue
• MC Dropout: Keeping dropout enabled during testing, obtaining prediction distributions via multiple forward passes with almost zero additional cost
• Quantile Regression: Using the Pinball loss function to estimate prediction intervals without assuming a noise distribution
• Deep Ensembles: Training multiple independent models, using prediction variance to estimate uncertainty; excellent performance but high computational cost
• SNGP (Spectral Normalized Gaussian Process): Combining spectral normalization and Gaussian processes to obtain reliable uncertainty estimates while maintaining single-model efficiency

Classification tasks face unique challenges, especially since the Softmax function often produces overconfident predictions. This section explores:

• Softmax Overconfidence Problem: Why models can give 99% "confidence" even for misclassifications
• Temperature Scaling: A simple and effective post-processing calibration technique that improves confidence without modifying the model
• Application of MC Dropout in Classification: How to identify uncertain samples via prediction entropy and variance
• SNGP for Classification: More reliable prediction probabilities compared to Softmax
• Out-of-Distribution (OOD) Detection: How models should identify and reject predictions when inputs are completely different from training data

Theory needs to be put into practice. This section uses the EuroSAT satellite image dataset to demonstrate real-world UQ challenges:

• Using real convolutional neural networks to process satellite images
• Introducing distribution shifts via synthetic clouds to simulate data drift in real scenarios
• Distinguishing between aleatoric uncertainty (data inherent noise) and epistemic uncertainty (model knowledge gaps)
• Demonstrating how uncertainty correctly increases when input data quality degrades

Finally, an open-source tool is introduced to make UQ more than just a research concept:

• Lightning-UQ-Box: A complete UQ framework built on PyTorch Lightning
• Provides plug-and-play APIs supporting regression, classification, and segmentation tasks
• Built-in all methods introduced in the workshop, which can be directly copied and pasted into production code
• Automated training and evaluation workflows, lowering engineering barriers