Durée
30h Th, 60h Proj.
Nombre de crédits
| Master en science des données, à finalité spécialisée | 5 crédits | |||
| Master : ingénieur civil en science des données, à finalité spécialisée | 5 crédits |
Enseignant
Langue(s) de l'unité d'enseignement
Langue anglaise
Organisation et évaluation
Enseignement au premier quadrimestre, examen en janvier
Horaire
Unités d'enseignement prérequises et corequises
Les unités prérequises ou corequises sont présentées au sein de chaque programme
Contenus de l'unité d'enseignement
This course is about the loop that data analysis actually runs in: build a model, use it to analyse data, work out where it succeeds and fails, revise it, repeat. The emphasis is on the scientific method, with Bayesian modelling as the way of practising it.
You will learn to look at data before modelling it, to visualize it in ways that reveal rather than decorate, to write down a probabilistic model of how it was generated, to compute with that model (expectation-maximization, variational inference, MCMC), and then to criticize it: does it reproduce what we observe, and how would we know if it didn't?
Teaching is done with live coding. Derivations happen at the blackboard, then the same thing is implemented and run in a notebook in front of you. Readings are short papers from the primary literature rather than textbook chapters.
The lectures follow each step of the loop (subject to change):
- Lecture 1: Build, compute, critique, repeat
- Lecture 2: Data and exploratory analysis
- Lecture 3: Visualization
- Lecture 4: Latent variable models
- Lecture 5: State-space models
- Lecture 6: Markov chain Monte Carlo
- Lecture 7: Expectation-maximization
- Lecture 8: Variational inference
- Lecture 9: Model criticism and validation
- Lecture 10: Wrap-up case study
The course assumes you can program comfortably and are not intimidated by probability. It does not assume you have done Bayesian inference before.
Acquis d'apprentissage (objectifs d'apprentissage) de l'unité d'enseignement
At the end of the course, you will be able to take a real dataset and a real question and go the whole way: explore the data, propose a probabilistic model, fit it with an appropriate inference method, check whether it holds up, and state what it does and does not support. The case study project is where you do this end to end, on a dataset and a question given to you.
Savoirs et compétences prérequis
Programming experience. Probability and statistics. Elements of artificial intelligence.
Activités d'apprentissage prévues et méthodes d'enseignement
- Lectures, with live coding sessions
- Reading assignments
- A case study project
Mode d'enseignement (présentiel, à distance, hybride)
Cours donné exclusivement en présentiel
Informations complémentaires:
Lectures are taught face-to-face. The case study project is carried out remotely.
Supports de cours, lectures obligatoires ou recommandées
Autre(s) site(s) utilisé(s) pour les supports de cours
- Github (github.com/glouppe/dats0001-foundations-of-data-science)
Informations complémentaires:
Materials are made publicly available on GitHub during the semester.
Modalités d'évaluation et critères
Examen(s) en session
Toutes sessions confondues
- En présentiel
évaluation orale
Travail à rendre - rapport
Informations complémentaires:
The evaluation has two components:
- Case study project: 40%
- Oral exam on the theory: 60%
The project is submitted as a notebook and is not graded on the notebook alone. It also includes an oral defense, where you answer direct questions on your own code, justify your modelling choices, and may be asked to modify or extend your analysis.
A minimum of 8/20 is required for each component. Below that, the lowest of the two grades becomes the final grade.
Use of AI tools
Using AI assistants is allowed, without restriction. The rule is simple: everything you submit, you must be able to explain. Submitting an analysis you do not understand, because an assistant produced it, will be strongly penalized. This is why the evaluation is built the way it is: the exam is oral, and the questions on your notebook are asked with you in the room. A model can write your analysis, but it cannot sit the exam, justify why you chose that likelihood over another, or extend your model on request.
Stage(s)
Remarques organisationnelles et modifications principales apportées au cours
The course hub is available at https://github.com/glouppe/dats0001-foundations-of-data-science
Contacts
Gilles Louppe (g.louppe@uliege.be)