Computational modeling of cortical pathways involved in action execution and action observation

Brain imaging studies in macaque monkeys have recently shown that the observation and execution of specific types of grasp actions activate the same regions in the parietal, primary motor and somatosensory lobes. This extended overlapping pathway of activations provides new insights on how primates are able to learn during observation. It suggests that an observed behavior is recognized by simulating it using the circuitry developed for action execution. In the present paper we consider how learning via observation can be implemented in an artificial agent based on the above overlapping pathway of activations. We demonstrate that the circuitry developed for action execution can be activated during observation, if the agent is able to perform action association, i.e. to relate its own actions with the ones of the demonstrator. Following this intuition, a computational model of observation/execution of grasp actions is developed and used in experiments to study its properties and learning capacities. Results show that the agent is able to associate novel objects with known behaviors only by observation. Model investigation after training reveals that, during observation, not only the same regions, but also the same neural representations are activated, implying that an observed action is understood by employing the same neural codes used for its execution.