Skip to content

Classes (from 2026–27)

New five-year accreditation

This page describes the new curriculum starting fall 2026 under the new quinquennal accreditation. For the previous program (up to 2025–26), see Classes up to 2025–26.

You may use this guide to select your UE (Unité d'Enseignement), whether you are enrolled in the [AI] track or another track of the Informatics Master's program.

All [AI] courses are open to students from other tracks; however, please ensure that you meet the required prerequisites.

If in doubt, contact the course instructors or the AI Master's program secretary or coordinators.

You can also catch up during the summer by taking one of our online crash courses.

M1 Classes

P1 — First period (September–October)

Applied Statistics Domain

This course introduces the necessary formalism to solve computational problems with statistical reasoning. The goal is to provide students with the necessary skills to employ statistical modeling in a computational framework to tackle real world problems.

Prior knowledge: basic linear algebra.

Acquired skills:

  1. Give a statistical description of a dataset.
  2. Leverage probabilistic modeling to perform statistical inference.
  3. Analyze the mathematical properties of a statistical estimator.
  4. Employ bootstrapping.
  5. Perform hypothesis testing.

References: All of Statistics: A Concise Course in Statistical Inference — Larry Wasserman (PDF)

Hands-on Machine Learning with Scikit-learn Domain

A practical oriented class, where students apply ML techniques to simple illustrative examples and then to tackle competitive challenges. It will start with an introduction to present (refresh) the ML landscape. Classes will then be articulated to successively focus on the major concepts of practical ML.

Outline:

  1. Introduction/refresher on ML
  2. Working with real data
  3. Discover and visualize the data to gain insights
  4. Prepare the data for processing
  5. Select and train models
  6. Fine-tune models

Recommended reading:

  • Géron (2019) Hands-on Machine Learning with Scikit-Learn, Keras, and Tensorflow
  • VanderPlas (2017) Python Data Science Handbook
Mathematics for Data Science Opening

This class aims at teaching/reminding mathematical basis useful in data science:

  1. Vector spaces, linear transformations
  2. Matrices, linear systems
  3. Trace, determinant
  4. Norms, orthogonality
  5. Eigenvalues, singular value decomposition
  6. Tensors (notions), multivariable calculus

The course is evaluated by a written exam.

References:

Foundations of Agent-based Systems Opening · DKAI

P2 — Second period (November–December)

Machine Learning Basics 1 Domain

This course is algorithms-oriented: we sketch the great principles of ML at first, and then focus on how algorithms work in practice, including all necessary mathematical aspects. They are the basic building blocks of more advanced algorithms.

  1. Gradient Descent, Linear Regression from scratch
  2. Classification with a single layer Perceptron, from scratch. Geometrical interpretation, SGD/mini-batch learning. Discussion on the choice of the loss function or activation functions. OVR multi-class scheme.
  3. Overfitting, train/validation/test split, K-fold CV, regularization (L2, L1)
  4. MAP, Bayesian interpretation of Ridge Regression or Lasso
  5. Feature maps ("Kernel trick"), PCA (from scratch, seen as variance maximization), PCA as pre-processing (dimensional reduction)
  6. Kernels, Kernelized perceptron, SVM in the separable case

Acquired skills:

  • Know the basics of ML vocabulary
  • Make good habits, understand the standard pipeline
  • Know standard supervised, shallow-ML algorithms (be able to write their pseudo-code, explain their functioning)
  • Be able to code an algorithm by reading its documentation

Recommended reading:

  • Bishop (2006) Pattern Recognition and Machine Learning (available online for free)
  • Géron Hands-on Machine Learning with Scikit-Learn and TensorFlow — also in French: Introduction au Machine Learning
Optimization Domain

This Optimization for Machine Learning course aims at studying the mathematical and computational constructions and properties of key optimization algorithms in different situations (continuous/discrete, constrained/unconstrained, convex/non-convex, deterministic, and stochastic problems) with use case illustrations in machine learning.

  1. Introductions and background (convexity, differentiability, optimality conditions, convergence rates…)
  2. Continuous optimization (first order methods: gradient methods, linear search, acceleration)
  3. Continuous optimization (second order methods: Newton methods including Quasi-Newton, secant, IRLS)
  4. Constrained optimization (equality and inequality constraints, duality/Lagrangian, KKT optimality conditions, linear programming, GD for a constrained problem)
  5. Non-convex, stochastic optimization (the EM algorithm, stochastic gradient, stochastic EM)

Readings:

  • Convex Optimization — Boyd and Vandenberghe
  • Numerical Optimization — Nocedal and Wright
  • Optimization — Kenneth Lange

Course page

Data Camp Domain Intensive week

Full-time intensive week (2–6 November 2026).

This course aims to learn the practical tools for data science and how to frame and solve data science problems.

  1. Data wrangling
  2. The scikit-learn API and missing values
  3. Metrics and unbalanced data
  4. Dealing with complex data
  5. Ensemble methods and hyperparameter optimization

Full syllabus on GitHub

History & Societal Impacts of AI Domain / Soft skill

This course aims to give students an overview of the development of artificial intelligence from its beginnings to its current state. Students will learn about key events, major ideas, and ethical issues in AI, helping them understand how today's AI tools came to be and their impact on society.

  • Introduction to AI and Its History — Overview of the course, discussion on what AI is, its current issues, and a brief look at the history of AI development.
  • Philosophical Foundations of AI — Introduction to key philosophical ideas in AI, including the Turing Test, John Searle's Chinese Room argument, and Leibniz's thoughts on computation.
  • Major Periods in AI Development — Overview of important times in AI history, focusing on the growth and decline of interest and funding in AI research (known as AI Summer and Winter).
  • Development of AI Techniques — Exploration of different methods used in AI, the link between linguistics and AI, and how gaming technology has influenced deep learning.
  • Machine Creativity and Future Predictions — Discussion on how AI can create art, music, and literature, and what the future might hold for AI, including the idea of the singularity.
  • Ethics and Future Challenges in AI — Examination of ethical issues, political considerations, and future challenges in AI, focusing on the responsibilities of AI researchers and the impact on society.

Ecampus

High Performance Computing on multicore architectures Opening · DIPAQ
Constraint Programming: Theory and Practice Opening · DKAI

P3 — Third period (January–February)

Machine Learning Basics 2 Domain

This course aims at mastering the core concept of algorithmic design in ML, from an optimization or a probabilistic point-of-view, using supervised and unsupervised algorithms.

  1. Regression/classification seen in optimization and probabilistic frameworks, implication on batch and stochastic gradient descent
  2. Learning theory and Vapnick-Chervonenkis dimension
  3. Evaluating performances of ML algorithms in different contexts (imbalanced, small-sized, etc.)
  4. Probabilistic framework for machine learning: Discriminative vs Generative learning, Empirical Risk Minimization, Risk Decomposition, Bias-Variance Tradeoff; MLE, MLE and OLS in regression, MLE and IRLS in softmax classification
  5. Unsupervised Learning and Clustering: K-means, Mixture Models, EM algorithms
  6. Unsupervised Learning and Dimensionality reduction: PCA, Probabilistic PCA & EM, ICA

Recommended reading:

  • James, Witten, Hastie & Tibshirani (2013) An Introduction to Statistical Learning
  • Hart, Stork & Duda (2000) Pattern Classification
  • Cornuéjols & Miclet (2011) Apprentissage artificiel: concepts et algorithmes
Hands-on Natural Language Processing (2 groups) Domain
  1. Introduction to NLP
  2. Basic concepts in NLP, tokenization, lemmatisation, POS tagging, ...
  3. Lexical semantics, word sense disambiguation
  4. Syntax and interpretations
  5. Parsing

Recommended reading: Natural Language Processing with Python

Creation of an AI Challenge Domain

This course is designed to bridge theory (lectures) and practice (TPs) by guiding students through the end-to-end creation and resolution of an AI challenge, using pre-formatted real datasets and the open-source platform Codabench.

  1. Designing AI Challenges
  2. Evaluation and Metrics
  3. Baselines: Core ML algorithms
  4. Baselines: Image Classification
  5. Optimizing ML Solutions
  6. Interpreting Results & Presenting Impact
  7. Project presentations

Full syllabus

Deliverables:

  • A Codabench challenge (including: website, GitHub repo, starting kit with a Python notebook)
  • A challenge/benchmark leaderboard with submissions from classmates
  • A written report (with conference publication potential if high quality)
  • An oral project presentation

P4 — Fourth period (March–April)

Deep Learning Domain

The aim of this course is to introduce the Deep-Learning framework. It will cover fundamental models such as the multilayer perceptron through to the most recent deep learning architectures (CNN, VAE, ...). In addition, the course will cover the different approaches to train these neural networks, with lectures dedicated to backpropagation algorithms and optimization methods based on gradient descent.

  1. Introduction to Neural Networks and the MLP model
  2. MLP and Gradient Descent algorithm
  3. Backpropagation algorithm and optimization methods
  4. Create your neural network with PyTorch
  5. Neural Networks architectures (CNN, AE, ...)
  6. Generative approaches (VAE, GAN, Denoising Diffusion models)

Acquired skills:

  • Broad view of NN architectures
  • Training and evaluating NN
  • Use of PyTorch
  • Implementing backpropagation algorithm

Recommended reading: Deep Learning Book

Speech and Language Processing Domain

Various talks on the following subjects:

  1. From linguistics to NLP
  2. Treebanks and oral syntax
  3. NLP, semantics, multi-word expressions
  4. Speech-Audio processing + ASR + practical work
  5. Emotion detection
  6. Speech interaction

Keywords: ongoing NLP research at LISN

Signal Processing Domain
  1. Introduction & Fourier analysis
  2. Filtering
  3. Random signals
  4. Time-Frequency analysis
  5. Time-Scale analysis
  6. Introduction to linear inverse problems

Keywords: spectral analysis of time signals and images, convolution and filtering, time-frequency and wavelets analysis, denoising, sparse coding for inverse problem resolution

Course page

2026–27 only: this course is also offered as Signal Processing M2 (P5, Friday morning), an alternative to From Symbolic to Neurosymbolic AI for M2 students.

Trustworthy AI: Fairness, Bias, and Responsibility Opening · DKAI

Instructor: Adrian Popescu — adrian.popescu@cea.fr

When and where: Tuesday 13:30–17:00, PUIO — D101.

This class explores the notions of ethics, bias, explainability, interpretability, privacy, and reliability in AI systems.

Acquired skills: recognize situations in which fairness issues are involved; take appropriate preventive or corrective actions to collect good data and to obtain predictive models that make fair decisions.

Prior knowledge: basic machine learning and statistics courses.

Evaluation: continuous assessment 70% + project 30%.

Course material on eCampus

Interactive Machine Learning Opening · HCI

Projects, internships, and other M1 modules

  • TER — Travaux d'Études et de Recherche / research training (6 ECTS, obligatory), runs across P3–P4
  • Stage (internship) or école thématique (summer school/seminars) (6 ECTS)
  • Soft-skill choices (e.g. Innovation et Entreprenariat, Impacts Environnementaux du numérique)
  • UE libre / Anglais / Français langue étrangère

For more details on internships and TER, see M1 Internships & TER. For the weekly schedule, see the Weekly Agenda.

M2 Classes (9 UE propres + 1 soft skill = 30 ECTS)

P5 — Fifth period (September–October)

Data Readiness and Evaluation in AI Systems Domain

9 hours of lectures + 12 hours of project work, in 6 sessions (replaces the former Frugal AI course).

This course addresses the preparation, documentation, and evaluation of the data needed to build reliable AI systems. Data is treated as a central object of the AI life cycle — collection, access, quality, representativeness, diversity, annotation, bias, and the fit between the available data and the target task. The goal is to show that the performance of an AI system depends as much on how its data are constituted and evaluated as on the choice of model.

  1. The role of data in AI — data readiness as a prerequisite for any AI project: relevance, quality, coverage, representativeness, and task/domain fit; cases where data are scarce, partial, costly to collect, subject to protection constraints, or only available sequentially. Presentation of the mini-projects.
  2. Collected and generated data — collection and generation as corpus-building methods; selection, framing, coverage, obsolescence, and representation biases; data documentation (collection conditions, populations covered, intended and prohibited uses).
  3. Annotation, label quality, and data reliability — crowdsourcing, collaborative and automated annotation, label aggregation, the notion of ground truth, annotator disagreement, subjective tasks. Labels as the outcome of a situated human process that must be documented and evaluated.
  4. Quantitative evaluation, experimental protocols, and the limits of metrics — how global metrics can hide differences between classes, groups, domains, time periods, or example types; baselines, test sets, representativeness, statistical robustness, external validity, and train/test contamination.
  5. Data audit and critical synthesis — identifying the strengths, weaknesses, and risks of a dataset or experimental protocol; distinguishing what can be claimed from the available data, what remains uncertain, and what would require further investigation.
  6. Project defenses in front of a jury of domain experts.

Practical work: group mini-projects on open research questions, run in parallel with the lectures (about 2 hours per session from session 2 to 5). Students analyse a concrete data-curation problem (availability, quality) or propose innovative quantitative or qualitative evaluation strategies, then discuss the limits of their protocol.

Keywords: data readiness, data documentation, annotation, bias, evaluation protocols, data audit, R&D methodology

NLP Today Domain

Instructors: Thomas Gerald, Sahar Ghannay

A tour of natural language processing from its early methods to current large language models, with the practical vocabulary and techniques in between.

Content for 2026–27, subject to change:

  1. Introduction, NLP Yesterday (Thomas Gerald)
  2. Tokenization, Stemming, Lemmatization (Thomas Gerald)
  3. Semantics and Word Embeddings (Sahar Ghannay)
  4. Open Domain vs. Specialty Domain (Sahar Ghannay)
  5. Language models and Transformer Architectures (Thomas Gerald)
  6. NLP Tomorrow (Thomas Gerald)

Keywords: text processing, word representations, neural networks for texts, language models

Course page (being updated for 2026–27)

Career Seminar Soft skill

When: P5, Friday morning — in parallel with Signal Processing M2.

Information: Vanessa Peña-Araya — vanessa.pena-araya@inria.fr

A seminar about what comes after the master. It is not only about preparing for a PhD: it covers career possibilities broadly — doctoral studies and academic research, research in industry, engineering and data-science positions, and how to get there. Speakers come to talk about their own paths.

Counts as your soft-skill UE (3 ECTS).

P6 — Sixth period (November–December)

From Symbolic to Neurosymbolic AI Domain · DKAI

Instructor: Yue Ma (LISN) — ma@lisn.fr

Why does symbolic reasoning still belong in an AI curriculum? For two reasons, and the course pursues both. It extends what language models can do — executing constraints and queries, checking proofs and countermodels, producing auditable evidence — and it trains the human, to tell validity from mere plausibility, to construct counterexamples, and to explain why a conclusion follows. Formalization is treated as an AI capability and as a reasoning skill at the same time.

By the end of the course you should be able to:

  1. Classify different inference patterns.
  2. Refute an invalid argument with a countermodel.
  3. Translate simple claims into propositional and first-order logic.
  4. Audit and repair a natural-language-to-logic translation.

Prior knowledge: no experience in knowledge representation is required. Prior exposure to logic, the Semantic Web, or knowledge graphs is helpful.

Keywords: symbolic reasoning, propositional and first-order logic, validity, countermodels, formalization, neurosymbolic AI

2026–27 only: M2 students choose either this course or Signal Processing M2 (P5, Friday morning).

Deep Learning for NLP Domain
  1. Introduction, language models
  2. Convolutional neural nets
  3. Recurrent neural networks
  4. Attention mechanisms
  5. Word representations, Transformers
  6. Generative models

Keywords: machine learning, statistics, probability theory, Python

Reinforcement Learning Domain
  1. Introduction to Reinforcement Learning
  2. Markov Decision Processes
  3. Planning by Dynamic Programming
  4. Model-Free Prediction
  5. Model-Free Control
  6. Value Function Approximation
  7. Policy Gradient Methods

Keywords: machine learning, statistics, probability theory, Python

Information Retrieval 2 — Recent Advances Domain

Called "Information Retrieval 2" in 2026–27 because this year's M2 students already took a course of that name in M1. It is a different, more advanced course.

Instructor: Kim Gerdes · six Friday mornings, 6 November – 11 December 2026 · a reading seminar with implementations, in English.

The M1 spring course built a complete modern retrieval pipeline: IR metrics → BM25 and variants → bi-encoders and FAISS → cross-encoder reranking, RRF, doc2query, ColBERT → BEIR/MTEB → a mini-RAG. This seminar takes that pipeline as given and asks what has changed since 2023, and what the open problems are.

Each pair of students adopts one paper — a modern classic or a 2024–2026 paper — presents it, reproduces a key result at small scale, and runs the method on a shared testbed, so that the seminar as a whole produces one comparable results table.

  1. Kickoff (6 Nov) — IR in 45 minutes, patents in 15, how to read and present an IR paper, a model presentation in the seminar format, hands-on baselines
  2. Representations (13 Nov) — how embedding models are trained today, contrastive objectives, embedding geometry, learned sparse and multi-vector retrieval
  3. Domains (20 Nov) — scholarly and patent retrieval, domain adaptation, synthetic training data, multilingual retrieval
  4. LLMs in the loop (27 Nov) — LLM rerankers, query rewriting, reasoning-intensive retrieval, generative retrieval
  5. RAG and agents (4 Dec) — adaptive and structured RAG, long context vs. retrieval, search agents
  6. Trust (11 Dec) — evaluation with LLM judges, poisoning and bias, multimodal documents · final synthesis

Your workload: one presentation with your partner, one turn as discussant, and one question posted before each session. Nothing else is graded.

Newcomers are welcome. Students who took neither the M1 spring course nor a comparable IR course elsewhere are paired with a returning student, get a self-study package (about 15 hours of slides and notebooks) and a 45-minute recap in session 1.

Bidding closes 30 September. The paper menu and the bidding page are online now: rank 3–5 papers. Pairs and papers are assigned on 5 October.

Course page, paper menu and bidding

P7 — Seventh period (January–February)

Advanced Optimization and Automated Machine Learning Domain
  1. Introduction to Optimization
  2. Unconstrained Continuous Optimization
  3. Constraint Optimization
  4. Black Box Optimization
  5. Hyper-parameter Optimization
  6. Neural Architecture Search + Learning to Optimize & Meta-Learning

Keywords: optimization, black-box optimization, Bayesian optimization, neural architecture search, hyperparameter optimization, meta-learning, AutoML

Probabilistic Generative Models Domain

This module covers probabilistic and neural generative models, from theoretical foundations to modern applications in vision, language, and multimodal AI.

  1. Foundations of Generative Models — definitions, probability basics, explicit vs. implicit models, and first applications
  2. Graphical Models & HMMs — Bayesian and Markov networks, exact/approximate inference, and sequential models
  3. Neural Generative Models (Images) — variational autoencoders (VAEs), GANs, and diffusion models for image synthesis
  4. Text Generation with Transformers — Transformer architecture, attention, pre-training (BERT, GPT)
  5. Optimization & Advanced LLM — efficient adaptation of large models with techniques like LoRA
  6. Emerging Models — Liquid Neural Nets, HRM, state-space & quantum-inspired approaches

Keywords: notions of statistics and AI

Scientific Machine Learning Domain

ML is increasingly adopted as a useful tool in the exact sciences (Physics, Chemistry, etc), where abundant precise data are available or can be generated with simulators. Applications include drug design, material design, genetics, applied quantum mechanics, fluid mechanics, and many others.

  1. Introduction to Scientific Machine Learning — applications in Physics, Chemistry, and Biology
  2. Graph Neural Networks and Geometric Deep Learning
  3. NeuralODEs and PINNs
  4. Steerable and Equivariant Neural Networks
  5. Attention Mechanisms across Architectures (GAT, EquiFormer, ViT)

This course introduces paradigms relevant to Scientific ML such as GNNs, NeuralODEs, PINNs, and the general framework of Geometric Deep Learning, including steerable neural networks (e.g. rotation-equivariant nets). It also dives deeper into the attention mechanism (e.g. for Vision Transformers), since the attention block is now of widespread use across other architectures.

Keywords: Graph Neural Networks (GNNs), NeuralODEs, Physics-Informed Neural Networks (PINNs), Geometric Deep Learning, Transformers

Soft skills & internship

  • 1 soft skill of choice (3 ECTS) — for instance the Career Seminar (P5, Friday morning, see above), Innovation et Entreprenariat, Impacts Environnementaux du numérique, or Français langue étrangère
  • Stage (30 ECTS) — 5 to 6 month internship in a research lab or a company