Evènement pour le groupe BioInformatique et Visualisation

Date 2016-04-14  11:00-12:00
TitrePrefrontal cortex model for language acquisition as abstract sequence learning applied to human robot interaction, and some canary songs ... 
RésuméThe topic of this presentation is embedded in the general topic of symbolic sequence processing with artificial neural networks. Primates can learn complex sequences that can be represented in the form of categories, and even more complex hierarchical structures such as language. The prefrontal cortex plays a major role in these processes. We model part of the prefrontal cortex using the Reservoir Computing framework: random recurrent neural networks allowing for high-dimensional dynamics. The model we generated, was applied to language processing using grammatical constructions. Even without world semantic information, the model network is able to generalize to unknown sentence constructions and make online predictions of thematic roles (" Who did what to whom? ") concerning future words. We propose a link between the model dynamics and Evoked-Related Potentials (ERP) observed when unusual sentences are processed by humans. The model was successfully applied to Human-Robot interaction in order to enable human users to teach the robot natural language sentences. Recently we showed that the very same instance of the model (same random connections) could learn both English and French sentence constructions simultaneously. A short glimpse on a current study about the structure of canary songs will also be given. ** Relevant references: * Core description of the model: http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0052946 * Model applied to user corpus for robot commands processing both English and French at the same time: https://www.researchgate.net/publication/284691419_A_Recurrent_Neural_Network_for_Multiple_Language_Acquisition_Starting_with_English_and_French * Implementation of the model on the iCub and inverse model for language production http://journal.frontiersin.org/article/10.3389/fnbot.2014.00016/abstract ** Video of an application to Nao robot with 2 collaborators (on 3D vision and speech): http://bit.ly/humanoidly-speaking ** My last project: EchoRob (that I just finished in Hamburg): https://www2.informatik.uni-hamburg.de/WTM/projects/EchoRob/EchoRob.shtml 
OrateurXavier Hinaut 

Aucun document lié à cet événement.

Retour à l'index