Optimization in machine learning

Summary: This course covers recent advances in scalable algorithms for convex optimization, with a particular emphasis on training (linear) predictors via the empirical risk minimization paradigm. The material will be presented in a unified way wherever possible. Randomized, deterministic, primal, dual, accelerated, serial, parallel and distributed methods will be mentioned. The course will start in an unusual place: a concise yet powerful theory of randomized iterative methods for linear systems. While of an independent interest, this will highlight many of the algorithmic schemes and tools we shall encounter later in the course.

Teacher: Peter Richtarik Peter Richtárik is an Assistant Professor of Optimization at the University of Edinburgh. His research interests are in all areas of data science that intersect with optimization, including algorithms, machine learning, statistics, operations research, mathematics and high performance computing.

 Reinforcement learning

Summary: The course will cover the basic models and techniques of reinforcement learning (RL). We will begin by reviewing the Markov decision process (MDP) model used to formalize the interaction between a learning agent and an (unknown) dynamic environment. After introducing the dynamic programming techniques used to compute the exact optimal solution of an MDP known in advance, we will move to the actual learning problem where the MDP is unknown and we will introduce popular algorithms such as Q-learning and SARSA. This will lead to the analysis of two critical aspects of RL algorithms: how to trade off exploration and exploitation, and how to accurately approximate solutions. The core of the exploration-exploitation problem will be studied in the celebrated multi-armed bandit framework and its application to modern recommendation systems. Finally, few examples of approximate dynamic programming will be presented together with some guarantees on their performance. The hands-on session will focus on implementing multi-armed bandit algorithms applied to the problem of policy optimization and online RL for simple navigation problems.

Teacher: Alessandro Lazaric, Bruno Scherrer Alessandro Lazaric received his PhD from the Electronic and Informatics Department of Politecnico di Milano, under the supervision of Andrea Bonarini and Marcello Restelli. He is currently a Junior Researcher (CR1) at INRIA Lille - Nord Europe in the SequeL team led by Philippe Preux and Rémi Munos. Bruno Scherrer is a research scientist at INRIA in the project BIGS. He is a member of the Probability and Statistics Team at Institut Elie Cartan of Lorraine (IECL). His main research interests are stochastic optimal control, reinforcement learning, Markov decision processes, approximate dynamic programming, analysis of algorithms and stochastic processes.

Dictionary learning

Summary: In recent years, a large amount of multi-disciplinary research has been conducted on sparse models and their applications. In statistics and machine learning, the sparsity principle is used to perform model selection---that is, automatically selecting a simple model among a large collection of them. In signal processing, sparse coding consists of representing data with linear combinations of a few dictionary elements. Subsequently, the corresponding tools have been widely adopted by several scientific communities such as neuroscience, bioinformatics, or computer vision. The goal of this course is to offer a self-contained view of sparse modeling for visual recognition and image processing. More specifically, we focus on applications where the dictionary is learned and adapted to data, yielding a compact representation that has been successful in various contexts.

Teacher: Julien Mairal Julien Mairal is a research scientist at INRIA in the project LEAR. He was previously a postdoctoral researcher in the statistics department of the university of California, Berkeley. Before that, he did his PhD at INRIA in the project WILLOW under the supervision of Jean Ponce and Francis Bach. He is interested in machine learning, optimization, computer vision, statistical signal and image processing, and also have some interest in bio-informatics and neurosciences.

 Information Retrieval and Machine Learning

Summary: This course is an introduction to the intersection between Information Retrieval (IR) and Machine Learning (ML) models. ML has been at the basis of some IR tasks such as document ranking and relevance feedback. On the other hand IR poses new challenges to ML because of the peculiar nature of the context in which data are observed.  In this course, I will introduce first the tasks of IR and then the utilization of some ML techniques to address these tasks.

Teacher: Massimo Melucci Massimo Melucci completed a PhD in Computer Science in 1996. Since 2001 he has been Associate Professor in Computer Science at the Department of Information Engineering of the University of Padua, Italy. He has been the co-chair of ESSIR 2009, lecturer of RuSSIR 2013 and speaker of the ICTIR 2013 tutorial on IR and quantum theory.  He is on the Editorial Board of the Journal of IR and Associate Editor of Computer Science Review (Elsevier).  He has been area chair of SIGIR, AIRS, CIKM, co-chaired the PC of the 11th Symposium on String Processing and Information Retrieval and chaired the organisation of the 2004 Algorithmic Learning Theory conference, of the 2004 Symposium on String Processing and IR, and of the 2004 Discovery Science conference. He has been PC chair of the 5th and 6th Quantum Interaction symposia, general chair of the 1st Italian Information Retrieval workshop in 2010 and PC co-chair of the 2nd Italian Information Retrieval workshop in 2011.  His research interests are mainly in IR modeling. He is also currently investigating the intersection between IR and machine learning and the use of quantum mechanics in IR. He has been involved in EU and national research projects.