Phelma Formation 2022

Bayesian cognition : Patterns for perception, learning and action - WPMGCOB7

  • Number of hours

    • Lectures 9.0
    • Projects 0
    • Tutorials 9.0
    • Internship 0
    • Laboratory works 0


    ECTS 1.5


Both living organisms and robotic systems must face the same central difficulty: How to survive while being ignorant? How can they use an incomplete and uncertain model of their environment to perceive, infer, decide, learn and act efficiently?

Indeed, any model of a real phenomenon is incomplete: there are always some hidden variables, not taken into account in the model, that nevertheless influence the phenomenon. The effect of these hidden variables is that the model and the phenomenon never behave exactly alike. Uncertainty is the direct and unavoidable consequence of incompleteness. A model may not foresee exactly the future observations of a phenomenon as these observations are biased by the hidden variables. It may neither predict exactly the consequences of its decisions.

Probability theory, considered as an alternative to logic to model rational reasoning, is the perfect mathematical framework to face this difficult challenge. Learning is used in a first step to transform incompleteness into uncertainty, inference is then used to reason and take decisions based on the probability distributions constructed by learning. This so-called subjectivist approach to probability allows uncertain reasoning as complex and formal as the ones made using logic with exact knowledge.

The main objective is to provide and introduction to this approach, from theoretical foundations to practical algorithms, and from central nervous system models to industrial applications.

Contact Julien DIARD


Content summary:

  • Theoretical foundation and justifications: how are probabilities an alternative to logic as a model of rational reasoning?
  • Formalism and models: how to build a Bayesian model?
  • Algorithms and inference: how are performed inference and learning? (Here, the description stays at a general audience level: we do not delve into the detail of Baum-Welch learning, MCMC or variational methods, for instance, contrary to what would be done in an algorithmic class).
  • Detailed examples of Bayesian models used in life sciences, robotics, and industrial applications.
  • Application of the Bayesian formalism for model comparison and selection.


Note for English speaking students: class is given in French, but slides are in a French-English mix, and I can answer in English if you ask questions in English. I advise you to take a look at slides beforehand to have an idea of the subject matter if you feel following the class in French would be too difficult.


An exam (EXAM1 for the main session, EXAM2 for the second session) provides the evaluation, for students of the M2 "Cognitive Science". There is no midterm or homework. For PhD students, the exam is optional, and the course can validate "scientific formation" credits.

SESSION 1 (présentiel) : 100% examen écrit
SESSION 2 (présentiel) : 100% examen oral
SESSION 1 (à distance) : 100% examen écrit
SESSION 2 (à distance) : 100% examen oral

Additional Information

Course list
Curriculum->Double-Diploma Engineer/Master->Semester 9
Curriculum->Master->Semester 9


Deux références d’introduction, disponibles sur :

  • F. Colas, J. Diard, and P. Bessière. Common bayesian models for common cognitive issues. Acta Biotheoretica, 58(2-3):191–216, 2010
  • O. Lebeltel, P. Bessière, J. Diard, and E. Mazer. Bayesian robot programming. Autonomous Robots, 16(1) :49–79, 2004