Choisir la langue :

Learning in actionable universes

Institutional tag: 
Thematic tag(s): 

The many constraints posed by a recommending system (online learning, large response sets, context adaptation, frequent renewal) are quite close to those encountered in embedded and biological control systems, having to take many decisions in limited time in  non-stationary environments. Bridging the  bandit scheme with effective action selection in biological and real-world interacting devices is thus a promising avenue,  that should both help enlighten some aspects of biological decision, and develop more "natural" behaviors in man-machine interactions.

After presenting some recent results regarding sparse online learning in a contextual bandit setup, I will develop on a more general “actionnable” universe learning setup. In particular, I should present directions  toward (i) modelling action selection in the brain, and (ii) develop active sensors that mimic the saccadic system in human and animal vision.

Dates: 
Tuesday, May 16, 2017 - 14:00
Location: 
Inria, room A00
Speaker(s): 
Emmanuel Daucé
Affiliation(s): 
Ecole Centrale Marseille