[PS-2.8] A dynamic neural field model for learning sequential tasks by observation

Sousa, E. , Erlhagen, W. & Bicho, E.

University of Minho, Portugal.

Sequential learning plays a fundamental role in a large variety of everyday tasks. Very often, the learning is challenged by the fact that the sequential order of events is not fixed and/or a hierarchical grouping of successive events is necessarily involved. Being able to encode, represent and recall a sequential task is thus a ubiquitous facet of higher cognition. Our group at the University of Minho develops autonomous robots able to assists human users in their daily activity in a natural and intuitive manner. An important challenge is to endow the robot with the capacity to learn the sequential organization of new tasks from user?s demonstrations and feedback. Here we present a dynamic neural field model (DNF) for interactive sequential learning and evaluate its real-time performance in the robot in the context of an assembly task. The DNF model consists of two connected layers, u_past and u_present, in which self-stabilized activity of neural populations encode already achieved and still to be accomplished sub-goals of the assembly work. During task demonstration, a Hebbian learning dynamics establishes new synaptic connections between sufficiently active populations in the two layers. Since the time course of supra-threshold population activity in u_past is adaptable, the number of previous events that become linked to a current sub-goal can be controlled in a task dependent manner. During execution, the populations in u_past drive through the newly established links the representation of the succeeding assembly step. The focus of the robotics implementation is on learning from different users that may show individual preferences in the serial order of task execution. A second objective is to show the robot?s ability to group different assembly steps together in order to represent subtasks that have to be completed.