probabilistic-ai — ETH Zürich Probabilistic Artificial Intelligence (2022–2023)

Production-style implementations of selected coursework from the Probabilistic Artificial Intelligence course at ETH Zürich, packaged as a single Python project.

Scope

• Bayesian inference — log-posterior over competing generative models
• Gaussian process regression — large-scale GP under an asymmetric cost
• Bayesian neural networks & calibration — MC dropout, deep ensembles, SGLD, and Bayes-by-backprop, with reliability diagrams and OOD detection
• Safe / constrained Bayesian optimisation
• Reinforcement learning — policy-gradient agent on LunarLander
• SWA-Gaussian (SWAG) — diagonal and full covariance approximate posteriors

Repository structure

.
├── src/probabilistic_ai/           # Python package
│   ├── task0_bayesian_inference/
│   ├── task1_gp_regression/
│   ├── task2_bnn_uncertainty/
│   ├── task3_constrained_bo/
│   ├── task4_rl/
│   └── task5_swag/
├── tests/                          # automated tests
├── data/                           # dataset placeholders + docs
├── docs/                           # high-level technical docs
├── .github/                        # CI, templates, ownership metadata
├── pyproject.toml                  # Poetry project config
└── Makefile                        # common dev commands

Each task module includes a local README.md describing the problem, methodology, and execution details.

Quickstart

Prerequisites

• Python 3.10–3.12
• Poetry
• macOS system deps (for Box2D): brew install swig

Install

poetry install
poetry run pre-commit install

Common commands

make check      # lint + format-check + tests
make format     # auto-format source and tests
make test       # run pytest

Data

Datasets are intentionally excluded from version control. See data/README.md for required files and paths.

CI/CD

GitHub Actions runs Ruff lint, Ruff formatting check, and Pytest on every push and pull request. See .github/workflows/ci.yml.

License

MIT. See LICENSE.