Parcours Artificial Intelligence 2e année

Machine Learning is one of the key areas of Artificial Intelligence and it concerns the study and the development of quantitative models that enables a computer to perform tasks without being explicitly programmed to do them. Learning in this context is hence to recognize complex forms and to make intelligent decisions. Given all existing entries, the difficulty of this task lies in the fact that all possible decisions is usually very complex to enumerate. To get around that, machine learning algorithms are designed in order to gain knowledge on the problem to be addressed based on a limited set of observed data extracted from this problem. To illustrate this principle, consider the supervised learning task, where the prediction function, which infers a predicted output for a given input, is learned over a finite set of labeled training examples, where each instance of this set is a pair constituted of a vector characterizing an observation in a given vector space, and an associated desired response for that instance (also called desired output). After the training step, the function returned by the algorithm is sought to give predictions on new examples, which have not been used in the learning process, with the lowest probability of error. The underlying assumption in this case is that the examples are, in general, representative of the prediction problem on which the function will be applied. We expect that the learning algorithm produces a function that will have a good generalization performance and not the one that is able to perfectly reproduce the outputs associated to the training examples. Guarantees of learnability of this process were studied in the theory of machine learning largely initiated by Vladimir Vapnik. These guarantees are dependent on the size of the training set and the complexity of the class of functions where the algorithm searches for the prediction function. Emerging technologies, particularly those related to the development of Internet, reshaped the domain of machine learning with new learning frameworks that have been studied to better tackle the related problems. One of these frameworks concerns the problem of learning with partially labeled data, or semi-supervised learning, which development is motivated by the effort that has to be made to construct labeled training sets for some problems, while large amount of unlabeled data can be gathered easily for these problems. The inherent assumption, in this case, is that unlabeled data contain relevant information about the task that has to be solved, and that it is a natural idea to try to extract this information so as to provide the learning algorithm more evidence. From these facts were born a number of works that intended to use a small amount of labeled data simultaneously with a large amount of unlabeled data to learn a prediction function.

The intent of this course is to propose a broad introduction to the field of Machine Learning, including discussions of each of the major frameworks, supervised, unsupervised, semi-supervised and reinforcement learning.

Voir la page complète

This course is related to mathematical and statistical methods which are very used in supervised learning.

It contains two parts.

In the first part, we will focus on parametric modeling. Starting with the classical linear regression, we will describe several families of estimators that work when considering high-dimensional data, where the classical least square estimator does not work. Model selection and model assessment will particularly be described.

In the second part, we shall focus on nonparametric methods. We will present several tools and ingredients to predict the future value of a variable. We shall focus on methods for non parametric regression from independent to correlated training dataset. We shall also study some methods to avoid the overfitting in supervised learning.

This course will be followed by practical sessions with the R software.

Voir la page complète

Voir la page complète

Causality is at the core of our vision of the world and of the way we reason. It has long been recognized as an important concept and was already mentioned in the ancient Hindu scriptures: “Cause is the effect concealed, effect is the cause revealed”. Even Democritus famously proclaimed that he would rather discover a causal relation than be the king of presumably the wealthiest empire of his time. Nowadays, causality is seen as an ideal way to explain observed phenomena and to provide tools to reason on possible outcomes of interventions and what-if experiments, which are central to counterfactual reasoning, as ‘‘what if this patient had been given this particular treatment?’’

Voir la page complète

Statistical learning is about the construction and study of systems that can automatically learn from data. With the emergence of massive datasets commonly encountered today, the need for powerful machine learning is of acute importance. Examples of successful applications include effective web search, anti-spam software, computer vision, robotics, practical speech recognition, and a deeper understanding of the human genome. This course gives an introduction to this exciting field. In the first part, we will introduce basic techniques such as logistic regression, multilayer perceptrons, nearest neighbor approaches, both from a theoretical and methodological point of views. In the second part, we will focus on more advanced techniques such as kernel methods, which is a versatile tool to represent data, in combination with (un)supervised learning techniques that are agnostic to the type of data that is learned from. The learning techniques that will be covered include regression, classification, clustering and dimension reduction. We will cover both the theoretical underpinnings of kernels, as well as a series of kernels that are important in practical applications. Finally we will touch upon topics of active research, such as large-scale kernel methods and the use of kernel methods to develop theoretical foundations of deep learning models.

Voir la page complète

The objective of this course is to present different techniques to handle uncertainty in decision problems. These techniques will be illustrated on several applications e.g. inventory control, scheduling, energy, machine learning.

Syllabus : Introduction to uncertainty in optimization problems; Reminders (probability, dynamic programming, ...); Markov chains; Markov decision processes; Stochastic programming; Robust optimization

Voir la page complète

Voir la page complète

The automatic processing of languages, whether written or spoken, has always been an essential part of artificial intelligence. This domain has encouraged the emergence of new uses thanks to the arrival in the industrial field of many technologies from research (spell-checkers, speech synthesis, speech recognition, machine translation, …). In this course, we present the most recent advances and challenges for research. We will discuss discourse analysis whether written or spoken, text clarification, automatic speech transcription and automatic translation, in particular recent advances with neural models.

Information access and retrieval is now ubiquitous in everyday life through search engines, recommendation systems, or technological and commercial surveillance, in many application domains either general or specific like health for instance. In this course, we will cover Information retrieval basics, information retrieval evaluation, models for information retrieval, medical information retrieval, and deep learning for multimedia indexing and retrieval

Voir la page complète

This course provides an introduction to computer vision. It concerns techniques for constructing systems that observe and recognize objects,  scenes and activities. It provides training in tools and techniques and models for: the image formation process,  color and illumination, image signal processing, multi-scale image description, image analysis, object detection, recognition and tracking, motion capture, modeling and understanding,  image matching,   multi-camera systems, and 3D reconstruction and modeling.

Voir la page complète

Voir la page complète

A robot is a mechatronic system with perception, decision and action capabilities design to perform in an autonomous way different tasks in the real world. Whatever the robot (e.g. mobile robot, industrial arm, mobile manipulator) and the task that it has been assigned, the robot will have to move (move its whole body or a part of its body, e.g. arm, hand). Accordingly, motion autonomy is an essential skill for a robot. To achieve motion autonomy, it is required to solve a number of challenging problems in areas as diverse as sensor data processing, situation understanding, motion planning, obstacle avoidance and control. The purpose of the course is to present the main concepts, tools and techniques that Roboticists have developed in the past fifty years in order to address these challenges. The course has three parts that focus on different aspects: The first part is about robot state estimation and world modeling. It presents the most popular approaches to perform state estimation. The basic equations of the Bayes filter are derived first. Then, the Extended Kalman Filter is introduced. These methods are then used to explore the following fundamental estimation problems: 1) robot localization, 2) Simultaneous Localization and Mapping (SLAM), 3) cooperative localization, and 4) simultaneous localization and self-calibration. The structural properties of these problems are studied. In particular, it is shown how the computational complexity scales with the size of the state. Finally, more theoretical aspects related to estimation with special focus on state observability are discussed. The second part focuses on the decision-making aspects. Motion planning is addressed first in the seminal configuration space framework, the main configuration space-based motion planning techniques are reviewed. Then, to deal with the uncertainty of the real world and the discrepancy between the world and its model, reactive collision avoidance techniques are presented. Finally, motion safety is formally studied thanks to the Inevitable Collision State concept. The third part is an introduction to control theory for articulated robots. The objectives are to understand basic concepts about the kinematics and dynamics of articulated robots and basic control theory in order to approach classical control methods, as well as a few selected advanced topics. The kinematics of articulated robots is introduced first, covering advanced topics such as singularities, hierarchies of objectives, inequality constraints. A brief reminder about Newton, Euler and Lagrangian equations of motion as well as basic Lyapunov stability theory is also provided before discussing standard motion control schemes such as Proportional-Derivative, Computed Torque, Operational Space and Task Function approaches. Advanced topics such as space robots, biped robots, Viability theory and optimal control are also touched.

Voir la page complète

This is an advanced lecture in Computer Graphics. It is divided in the three topics Geometric Modeling, Animation and Rendering. In Geometric Modeling we focus on the theory behind subdivision curves and surfaces. We present the main results for proving the convergence of a subdivision scheme and the smoothness of its limit. We provide practical recipes for applying these theoretical results to subdivision schemes. We illustrate the results in a homework in which students program subdivision curves and visualize their smoothness. The animation part is composed of 12 lessons presented during 4 lectures. Each lesson focuses on one important concept in computer animation, one of the 12 principles of traditional animation, and one research paper. This includes Keyframe animation, Forward and inverse kinematics, Motion planning, Rigging and skinning, and Mass spring systems. The rendering part focus on how to display a picture on the screen, so that it looks pleasing for the viewer. The topics presented include Materials and BRDFs, Shadow and shadow maps, Global illumination techniques, Monte-carlo ray-tracing , Level-of-Detail techniques and Expressive rendering. The main rendering techniques are implemented by the students in three homeworks.

Voir la page complète

Voir la page complète

Interactive Information Visualization (InfoVis)

InfoVis is the study of interactive graphical representations of abstract data (e.g. graphs linking people in social networks, series of stock options values evolving over time).

Graphical representations are a powerfull way to leverage the human perceptual capabilities to allow the user to explore and make sense of abstract data, and also to expose findings and convey ideas.

But to be efficient, a visualization has to be designed using knowledge about the human visual perception, the characteristics of the data, the kind of task that will be performed on those data.

The aim of this course is to provide the keys, both theoretical and practical, to build usable and useful interactive visualizations.

program summary:

- foundations: human visual perception, graphical variables, data types, the visualization pipeline.

- linked data: tree and graph visualization

- tabular data: time series and spatial data visualization

- dealing with large data: aggregation, multiple views, interaction

- validating visualization: visualization tasks, evaluation

Voir la page complète

This course introduces the field of MAS, various theoretical aspects (agent architectures, reasoning, interactions, game theory, social choice, etc), as well as practical applications from recent research. The focus is mostly on agent-based social simulation, and how to integrate psychological aspects in agents (so-called “human factors”: emotions, biases…) to make them more human-like and realistic. Applications discussed include epidemics modelling, computational economy, crisis management, urban planning, serious games, etc. The practical part of the course comprises several tutorials with various agent-based modelling platforms (in particular GAMA and Netlogo), scientific papers discussions, and analysis and/or extension of existing models. 

Voir la page complète

Voir la page complète

Voir la page complète

Voir la page complète

This course presents advanced methods and technics for Operations Research.

Reminder :

Linear Programming, Dynamic Programming, MIP modelling and BB
Complexity (P, NP, Co-NP)

Advanced MIP :

formulation, cuts, bounds
applications
lagragian relaxation
column generation
Benders decomposition
Solvers

Constraint Programming

Heuristics

local search
approximation algorithms

Voir la page complète

In this course, we will introduce parallel programming paradigms to the students in the context of applied mathematics. The students will learn to identify the parallel pattern in numerical algorithm. The key components that the course will focus on are : efficiency, scalability, parallel pattern, comparison of parallel algorithms, operational intensity and emerging programming paradigm. Trough different lab assignments, the students will apply the concepts of efficient parallel programming using Graphic Processing Unit. In the final project, the students will have the possibility to parallelize one of their own numerical application developed in a previous course.

1. Introduction to parallelism
2. Introduction to general context of parallelism
3. Models of parallel programming
4. Description of various model of parallelism
5. Paradigm of parallelism
6. Templates of parallelism
7. Parallel architectures
8. Programming tools: Cuda

Voir la page complète

http://mosig.imag.fr/MainEn/Projects

Voir la page complète