Differences

This shows you the differences between two versions of the page.

--- curriculum [2025/07/24 09:30] – respai-vic
+++ curriculum [2025/10/14 13:32] (current) – [Period 1] respai-vic
@@ Line 70: / Line 70: @@
 **APM_53440_EP - Advanced unsupervised learning (24h, 2 ECTS), Pierre Latouche(UCA)** (contact: pierre.latouche@uca.fr)
-> ABSTRACT TO BE ADDED
+This course aims at presenting advanced models and methods from computational statistics and machine learning for unsupervised learning, through the context of directed graphical models. Both the frequentist and Bayesian frameworks will be covered. In particular, we will study clustering methods and we will focus on mixture models and on the expectation maximisation algorithm for inference. The standard model selection criterion will be derived and we will illustrate their use to estimate the model complexity. After having given the theory of directed graphical models, we will show on a series of examples how the classical models from machine learning can be characterized in such a context. We will show how all the standard loss functions in machine learning can be linked to specific directed graphical models, giving rise to new ways to think about problems and to solve them. We will particularly insist on the interest of relying on directed graphical models to evaluate the complexity of the inference task for various models. The variational framework will be presented along with the variational EM and variational Bayes EM algorithms. The inference task of the stochastic block model, with a two-to-one dependency, will be described. The second part of the course will be about the use of deep neural networks in directed graphical models to obtain the so-called deep graphical models. Variational autoencoders and GAN we will be presented in such a context along with the corresponding inference strategies. We will finally show on a given example how the main model for social network analysis can be rewritten and extented with deep graphical models. We will point out the modifications in terms of directed graphical models, inference, and applications. In particular, we will describe the use of graph neural networks. Lectures will be done through slides and proofs on the blackboard. In labs, we will implement and test all the approaches seen in the lectures.
-**APM_53441_EP - Learning with tabular data (24h, 2 ECTS), Marine Le Morvan** (contact: marine.le-morvan@polytechnique.edu)
+**APM_53441_EP - From Boosting to Foundation Models: learning with Tabular Data (24h, 2 ECTS), Marine Le Morvan** (contact: marine.le-morvan@polytechnique.edu)
-> ABSTRACT TO BE ADDED
+> This course offers a modern overview of statistical learning with tabular data, from classical tree-based models to emerging deep and foundation models. We will review the foundations of gradient boosting methods and their scalable implementations, cover recent deep learning models tailored for tabular data, and introduce tabular foundation models. In particular, we will discuss the limitations of LLMs on structured data, introduce the concept of in-context learning, and provide an in-depth understanding of novel tabular foundation models, including their architecture and pretraining strategies. The course will also address key practical challenges in real-world datasets and applications, such as encoding heterogeneous feature types (categorical, numerical, temporal, textual), strategies for handling missing data, and methods for evaluating and calibrating predictive uncertainty.
@@ Line 88: / Line 92: @@
 **CSC_53433_EP - Creative & Generative models in Computer Graphics (24h, 2 ECTS),  Marie-Paule Cani (EP), Julien Pettré (Inria)** (contact: marie-paule.cani@polytechnique.edu)
-> This course presents the AI-related methods developed in Computer Graphics to create or generate individual 3D shapes, animated landscapes and humanoid motion. We first introduce user-centered Creative AI, i.e. smart 3D models - either based on knowledge or trained from examples, designed to help users creating and controlling 3D shapes and environments. Second, we focus on the use of AI in Character Animation, from early motion planning and control methods to deep reinforcement learning solutions. These methods result into 3D character models able to navigate alone or within crowds, and to interact with their environment. The lab sessions are held on Unity, based on C#.
+> This course presents the techniques for content creation in 3D Computer Graphics, from 3D shapes to animated landscapes and character motion, with a focus on their link with AI. We first introduce user-centered Creative AI (also called Expressive modeling methods), i.e. smart 3D models - either based on knowledge or trained from examples, designed to help users  create/generate the 3D shapes their have in mind, and their organisation into complex, static or animated scenes. Second, we study the use of AI in Character Animation, from early motion planning and control methods to deep reinforcement learning solutions for generating the gaits of physically-based characters. These methods result into 3D agents able to navigate alone or within crowds, and to interact with their environment. The lab sessions are held on Unity, based on C#.