MSc in Artificial Intelligence

This course will provide an overview of the field of Natural Language Processing.It will take a problem-centric approach, introducing increasingly complex problems, starting with fundamental building blocks such as language modelling, tagging, and parsing, and progressing to more advanced areas such as opinion mining, machine translation, question & answering, and dialogue. While critical historical methods will be mentioned (and studied if still relevant), a focus will be on current state-of-the-art, including many recent advances in neural network training and novel architectures.

Computer vision is an interdisciplinary field that aims to enable the automatic processing and understanding of digital images. These images can be captured by video cameras, radars, satellites, or even very specialised sensors such as those used for medical purposes. This course covers both theoretical and practical aspects of computer vision. It has been designed to provide an introduction to computer vision, including fundamentals on image processing, computational photography, multiview geometry, stereo, tracking, segmentation and object detection. It will expose students to several real-world applications that are important to our daily lives. The course will focus solely on classical computer vision techniques, providing a detailed overview of classical methods, but it does not cover deep learning-based methods.

Reinforcement Learning (RL) methods lie at the intersection of several fields, including machine learning, probability theory, control theory, and game theory. They study fundamental trade-offs between exploitation and exploration to optimise future rewards in an unknown environment. Recently, they were successfully applied to many applications, such as complex games (Go or Atari), robot control, and expert selection, leading, in many cases, to superhuman performance in those contexts. This class will cover the necessary tools (both theoretical and practical) to understand how such a breakthrough was possible.

Deep learning methods are now state-of-the-art across many machine learning tasks, yielding impressive results. Nevertheless, neural networks are still poorly understood, neural network training is still tricky, and the results are black boxes lacking explanations. Given the societal impact of machine learning techniques today (used for medical assistance, hiring processes, and bank loans…), it is crucial to make their decisions explainable or to offer guarantees. Besides, real-world problems usually do not fit the standard assumptions or frameworks of the most famous academic work (data quantity and quality, availability of expert knowledge…). This course aims to provide insights and tools to address these practical aspects, grounded in mathematical concepts.

Networks (or graphs) have become ubiquitous as data from diverse disciplines can naturally be mapped to graph structures. Social networks, such as academic collaboration networks and interaction networks over online social networking applications are used to represent and model social ties among individuals. Information networks, including the hyperlink structure of the Web and blog networks, have become crucial media for information dissemination, offering an effective way to represent and navigate content. A plethora of technological networks, including the Internet, power grids, telephone networks, and road networks, are essential to everyday life. The problem of efficiently and effectively extracting meaningful information from large-scale graph data has become crucial and challenging across several essential applications. Towards this end, graph mining and analysis methods constitute prominent tools. The goal of this course is to present recent, state-of-the-art techniques and algorithms for analysing, mining, and learning large-scale graph data, as well as their practical applications across various domains (e.g., the web, social networks, recommender systems).

Medical imaging technologies provide unparalleled means to study the structure and function of the human body in vivo. Interpretation of medical images is difficult because it requires accounting for three-dimensional, time-varying information across multiple image types. Artificial intelligence (AI) holds great promise for assisting in interpretation, and medical imaging is one of the areas where AI is expected to yield the most significant successes. In the past years, deep learning technologies have led to impressive advances in medical image processing and interpretation. This course covers both theoretical and practical aspects of deep learning for medical imaging. It covers the main tasks in medical image analysis (classification, segmentation, registration, and generative models). It presents state-of-the-art deep learning techniques alongside more traditional image processing and machine learning approaches. Examples of differenttypes of medical imaging applications (brain, cardiac…) will also be provided.

Ensemble methods are popular machine learning techniques that are powerful for both classification and regression. The idea is to build a global prediction model by combining the strengths of a set of simpler base models. The random Forests approach is an example that falls into this category by aggregating a collection of decision trees.

This course aims to study the fundamental notions and concepts of tree-based methods and ensemble learning. We will mainly introduce several methods, such as Classification and Regression Trees (CART), boosting trees, and Random Forests. Furthermore, we want to investigate and discuss recent challenges in interpretability and regularisation using different approaches [Strobl et al., 2007, 2008]. Moreover, the various methods will be applied to practical cases involving real data throughout the course.

Multi-agent systems (MAS) are widely used, especially for complex applications that require interaction among several entities. More specifically, they are used in applications where it is necessary to solve problems in a distributed manner (e.g., data processing) or in the design of distributed systems in which each component has some degree of autonomy (e.g., process control). Examples of applications include trading agents, drones, and smart grids. This course will begin with an introduction to agents and multi-agent systems. It will present the concepts that will lead to an understanding of what an agent is and how it can be built. Then, we will address a classical problem in SMAs: how to model and simulate a situation using the concept of agents.The idea is to provide a basis for understanding how agent-based simulations can be used as tools for understanding human societies and complex problems and situations. Besides, we will discuss how agents can communicate and interact to solve problems, and more specifically, to make noncentralized decisions. For this, we will rely on negotiation protocols based on argumentation theory, a process of constructing and evaluating arguments (positive and negative reasons/evidence) to resolve conflicts.

Maintenance is essential in engineering systems. In this course, we focus on the optimal planning of maintenance for complex systems. Different maintenance strategies are introduced with a focus on the performance indicators used to quantify their performance and optimization models used to obtain optimal planning. The students are supposed to learn fundamental theoretical results as well as tools to help them apply theoretical methods in practice.

Planning in AI focuses on automatically generating sequences of actions to reach a predefined goal. Actions are represented by operators that transform states, generally expressed in propositional logic. Several problems in industry or robotics can be formulated as planning issues. Different AI techniques and algorithms are used to solve planning problems. In classical planning, an agent has perfect knowledge of the states and the effects of the actions, which are deterministic. The relaxation of these assumptions leads to numerous extensions that use solving techniques based on diverse AI tools, such as heuristic search or Markov processes.

In this course, we will introduce the formulation of a planning problem and present the different AI techniques and algorithms for solving it.

Artificial Intelligence aims to solve problems automatically. In particular, Deep Learning enables achieving performance levels often exceeding human performance. However, in some application areas, users want to understand why an algorithm proposes a specific decision. This is especially true for human-centred domains, such as medicine and security, or for those affected by laws granting a right to explanation (the banking sector). For these reasons, the field of XAI (eXplainable Artificial Intelligence) has become very important in recent years, even though the issue has been around for several decades.This module aims to introduce you to tools that have emerged from recent or older research to help you understand the most opaque models and to familiarise you with interpretability and explainability issues. At the end of this course, the student will have acquired a general view of the XIA domain and its problems.

Decision-making is an indispensable task in engineering practices. In this course, we focus on the basic concepts, principles, and decision-making techniques using optimization models, especially in the presence of uncertainties.

The students are supposed to learn to model optimization problems involving risk and uncertainty using analytical models and off-the-shelf software. The course covers analytical models such as the basics of Convex Optimisation, Stochastic Programming, Risk Control in Stochastic Optimisation, Two-stage Stochastic Programming, Decomposition algorithms, Markov Decision Process, and Dynamic Programming. The course is hands-on. The emphasis will be on model formulation, practical problem-solving, implementation, and interpretation of results.

Artificial intelligence cannot be applied straightforwardly to finance for three essential reasons: data are extremely noisy, their distributions are usually far more non-Gaussian, and everything is nonstationary. These basic facts of life in finance, therefore, require special care, that is, specific methodologies and expertise.

This course aims at giving a practical approach to applying AI to finance: starting from fundamental concepts and objects, it then explores price formation in high-frequency data in limit order books, introduces modern methods of portfolio optimization, explains the good practices of trading strategy performance backtesting, and finally reviews what kind of DL/ML/RL approaches work in finance.

To complete the CentraleSupélec MSc in Artificial Intelligence, students must acquire professional experience (by means of a 4-6 month internship or a job in a position related to the program) and complete the related Master’s Thesis. This corresponds to a natural extension of the coursework carried out during the academic year