The MATHESIS project aims to explore fundamental aspects of social communication and adaptive behaviour, especially the process of assigning meaning to the actions of other subjects. This will be demonstrated by developing and validating artificial cognitive agents, able to acquire a repertory of motor actions by observational learning. Observational learning is understood here as the capacity to acquire an action strategy only through observation of other agents, without the experimentation needed in other learning procedures.
Our neurophysiological investigations will attempt to establish that the neural representations of action-execution, action-observation and action-recall overlap extensively within the cortex, as suggested by our preliminary results. A major implication of this, both for biological and for artificial agents, is that it should be possible to train their motor system by simple action-observation and action-recall. The representations, shared by both overt actions and by mentally simulated actions, may account for the traditional use of observational learning and mental training (e.g., the use of videotapes to train athletes).
These preliminary results motivate our study of neural network models and robotic cognitive agents, as they learn to control effectors via observation. In contrast to previous such models (based on the mirror neuron concept), which include action recognizing processes on a side path, we intend to use embodied agents to explore the emergence of action recognition in the controllers themselves. Moreover, we intend to examine if observation of action is an efficient means to parametrically adapt a small number of controllers guiding the behaviour of artificial agents.
Within MATHESIS we will assess the generality, scalability, accuracy and robustness of this cognitive architecture. Furthermore, we intend to establish the developmental stage at which observational learning can be used efficiently in infants and children.