🎓 Uncertainty Quantification in Deep Learning (Computer Vision mostly) - Master Thesis Repository

Welcome to the official code and resource repository for the master's thesis:

Title: Uncertainty Quantification in Deep Learning

Author: Nathaniel Cogneaux

Supervisor: Pr. Sungyong Baik

Institutions: University Paris Sciences & Lettres (France) & University of Padua (Italy)

Research Host: Hanyang University, Seoul, South Korea

Year: 2024

This repository accompanies the thesis, which introduces a novel post-hoc uncertainty quantification (UQ) technique that is efficient, model-agnostic, works in a single-pass and applicable to pre-trained deep neural networks without retraining.

Repository Contents

Master_Thesis_UQ/
│
├── dataloaders_and_metrics/        ← Data loaders and UQ metrics
│   ├── dataloaders.py              ← All dataset handling logic
│   └── metrics.py                  ← Metric definitions (ECE, NLL, etc.)
│
├── examples/                       ← Interactive usage examples
│   ├── Lenet.ipynb
│   ├── Wideresnet_28_10_cifar10.ipynb
│   └── Wideresnet_28_10_cifar100.ipynb
│
├── models/                         ← Model definitions
│   ├── lenet.py
│   ├── resnet.py
│   ├── vggnet.py
│   └── wide_resnet.py
│
├── multi_output_module/
│   └── multi_output_module.py     ← Our proposed meta-model for UQ
│
├── numerical_experiments/
│   ├── numerical_experiments.py   ← Script to reproduce main experiments
│   └── wandb_module_hyperparameters_tuning.py
│
├── Master_s_thesis_Nathaniel_Cogneaux.pdf   ← Full Thesis Document
└── dissertation_defense.pdf                ← Slides of the Defense Presentation

Project Highlights

✅ Post-Hoc UQ A simple, lightweight wrapper that estimates uncertainty without retraining.

✅ Model-Agnostic Compatible with any pre-trained model (LeNet, VGG, ResNet, WideResNet, etc.).

✅ Single-pass Provides multiple outputs with just one inference, no sampling is needed.

✅ Efficient & Scalable Achieves near SoTA results with drastically lower computational cost.

✅ Ready-to-Use Notebooks Step-by-step usage on MNIST, CIFAR-10, CIFAR-100.

✅ Benchmark Results Includes standard and corrupted datasets (CIFAR-10-C, CIFAR-100-C).

✅ Thesis & Defense Slides Theory fully explained and experimental insights.

Results demonstrate robust uncertainty quantification on:

• MNIST, FashionMNIST
• CIFAR-10, CIFAR-100
• Corrupted datasets: CIFAR-10-C, CIFAR-100-C

For a deep dive, please give a look at:

• Master_s_thesis_Nathaniel_Cogneaux.pdf, Full thesis (theory, method, results)
• dissertation_defense.pdf, Defense presentation slides

Quick Start

1. Clone the repo

   git clone https://github.com/NathanielCogneaux/Master_Thesis_UQ.git
   cd Master_Thesis_UQ

2. Create an environment

   python -m venv uq_env
   source uq_env/bin/activate

3. Install dependencies

   pip install -r requirements.txt

4. Try a notebook Launch Jupyter and open any notebook in examples/.

Reproducing Results

Run the main experiment script to regenerate results from the thesis:

python numerical_experiments/numerical_experiments.py

Or tune hyperparameters using:

python numerical_experiments/wandb_module_hyperparameters_tuning.py

All training/evaluation logic is managed through dataloaders_and_metrics/ and multi_output_module/.

📘 Citation

If you use this work in your research, please cite:

@mastersthesis{cogneaux2024masteruq,
  author       = {Nathaniel Cogneaux},
  title        = {Uncertainty Quantification in Deep Learning},
  school       = {University Paris Sciences & Lettres and University of Padua},
  year         = {2024},
  note         = {\url{https://github.com/NathanielCogneaux/Master_Thesis_UQ}},
}

Contact

For questions or collaborations:

• Open a GitHub Issue
• LinkedIn Profile