Probabilistic Machine Learning Seminar

About

This repository contains the code and experiments for my seminar thesis in the Probabilistic Machine Learning course (SoSe 2025).
The work focuses on exploring uncertainty in neural networks using partial Bayesian neural networks, with a particular emphasis on:

• Bayesian last-layer (Neural Linear Approximations)
  
• Comparison to fully deterministic neural networks
  
• Evaluation of uncertainty calibration and overconfidence on in-distribution and out-of-distribution (OOD) data

Core Topics

• Probabilistic modeling in neural networks
  
• Bayesian linear regression as output layer
  
• Predictive entropy and epistemic uncertainty
  
• OOD detection via uncertainty quantification
  
• Comparison to MAP-trained deterministic models

Project Structure

• src/ – Jupyter notebooks for experiments and visualizations

  • Simulation.ipynb – 1D toy regression to visualize predictive uncertainty
    
  • RealData.ipynb – Evaluation on real-world data
  • BayesianLLNN.py - Implementation of Bayesian Last Layer
  • baseNN.py - base deterministic NN

• models/ – Models

  • baseBayes.pth – base Neural Network for Classfication
  • basenn.pth - Standard ReLU neural network
    
  • best_lastlayer.pth - – Neural Linear Model (Bayesian output layer)

• Data/ – Data

  • Dataset of Diabetes.csv - Dataset in csv format

• results/ – Generated plots and saved evaluation results

• requirements.txt – Python dependencies

• README.md – This file

To set up the project environment and reproduce the results, follow these steps:

1. Clone the repository bash: git clone https://github.com/yourusername/probabilistic-seminar.git cd probabilistic-seminar
2. create virtual environment ``` # on Mac os: python -m venv venv source venv/bin/activate

# on Windows: venv\Scripts\activate 3. **install dependencies** pip install -r requirements.txt ```

1. Run experiments

For simluation study: Open and run src/Simulation.ipynb

For real-world classification: Open and run src/RealData.ipynb

Datasets Used

• Synthetic (toy) regression dataset — for visualizing predictive uncertainty
  
• Real-world dataset
  • Diabetes Diagnosis: available at Kaggle
  • In-distribution: known label classes
    
  • OOD: unrelated or random inputs for uncertainty evaluation

Goal of the Work

To evaluate whether a simple Bayesian approximation (only in the last layer) is sufficient to capture meaningful uncertainty in predictions — particularly in OOD settings — and to compare its behavior to a fully deterministic ReLU network.

LICENSE

This project is not open-source. All rights reserved. More infos found on License Unauthorized use, copying, or distribution is prohibited.