 | |
| ELEN0442-1 | Introduction to statistical training
 |
|
| Duration : | 30h Th, 30h Pr |
|
| Credits/ECTS : |
|
|
| Holder(s) : | Pierre Geurts, Louis Wehenkel |
|
| Language : | French language |
|
| Course contents : | Statistical learning denotes in a very general way the automatic construction of statistical models for a problem from a collection of data related to particular instances of this problem.
Its applications are multitudinous: medical diagnosis; econometrics and finance; bioinformatics; computer vision and automatic program synthesis; modeling, optimization, and control of complex processes.
This course first introduces the different types of statistical learning problems (supervised, unsupervised, semi-supervised, by reinforcement, active, ...) by illustrating them in the context of data exploration, automatic classification, and time-series forecast. Then the course exposes the principles underlying statistical learning methods (reformulation as optimization problems, bias/variance compromise, learnability, model validation).
Finally, the course elaborates on the statistical and algorithmic properties of the main classes of existing methods (decision trees, linear regression, neural networks, kernel based methods, ensemble methods). |
|
| Course objective : | Upon successful completion of the course, the students should master the basic principles of statistical learning. They will be able to read the scientific publications in the field and to use statistical learning methods to solve practical problems.
In particular, they will understand the theoretical bases of the field and know the essential properties of the main families of methods currently on the shelf. They will also be aware of the main research challenges in this field. |
|
| Prerequisites : | Statistical learning methods build on a very large range of methods and principles from classical calculus, euclidian geometry, probabilistic reasoning, statistics, numerical optimization, and combinatorial algorithms.
For this reason, this course is taught in a "problem based" fashion, so as to encourage the students to identify by themselves the need to acquire basics in these fields that they do not sufficiently well master. Except for the willingness to learn on the fly in these related fields, there are no other prerequisites for this course |
|
| Workshops : | Practical work is composed of two parts. In the first part the students become familiar with statistical learning by applying a representative sample of methods to a simple problem (simulated data) in a personal way.
In the second part, they carry out a research work by groups of two or three. They start from a problem formulation and a few scientific publications proposing solutions, in order to make a critical assessment of the proposed solutions. This assessment may be based either on related publications or on an implementation of the proposed algorithms. They present orally their work in front of the other students and the teachers. |
|
| Organization : | The course is organized during the 1st semester. On each Monday afternoon a theoretical course is given during 2 or 3 hours. Le practical works of the students completes these courses. |
|
| Written notes : | Course notes are composed of the handouts used during the theoretical course, some chapters of reference books in the field, and a number of founding papers in the field.
On-line notes : http://www.montefiore.ulg.ac.be/~lwh/AIA |
|
| Assessment : | Course evaluations are composed of three parts respectively related to the evaluation of the personal homework, the group project, and the oral exam. |
|
| Contacts : | Louis Wehenkel L.Wehenkel@ulg.ac.be
Pierre Geurts P.Geurts@ulg.ac.be |
|
| Remarks : | Website : http://www.montefiore.ulg.ac.be/~lwh/AIA |
|
|
| |
