Self-Supervised Laplace Approximation

This repository contains code to reproduce the experimental results for self-supervised laplace approximation across multiple settings, including:

• conjugate prior toy models,
• heteroscedastic regression, and
• real-world benchmarks (UCI datasets).

Getting started

Create and activate a virtual environment:

python -m venv .venv
source .venv/bin/activate  # Linux/macOS
# .venv\Scripts\activate  # Windows PowerShell

Install dependencies (Poetry):

pip install poetry
poetry install

Project structure

• experiments/ — runnable experiment suites and their outputs (tables/plots)
• src/ — core implementations (models, utilities, and SSLA/ASSLA code)
• scripts/ — convenience entry points to run experiments from the command line

Note on reproducibility

Some experiment scripts write their outputs into the corresponding experiments/**/ subfolders. Running an experiment may overwrite existing outputs. Where applicable, experiments are seeded so outputs should be reproducible.