A Mirror Neuron Inspired Model for Goal Inference
The human mirror neuron system (MNS) refers to brain regions becoming active both when an action is performed, and when a similar action is being observed. Several functions have been attributed to the MNS. It is argued that the MNS is responsible for goal inference, and could be involved in learning by imitation. Here we investigated an existing neural network model which exhibits mirror properties, the RecurrentNeuralNetworkwith Parametric Bias (RNNPB). However, this model fails to account for goal inference, when the body of the actor and observer are dissimilar. The RNNPB model is extended such that it is able to explain this general goal inference. This is achieved by ignoring bodily parameters as a direct input to the model, and instead use the perceived state of the world as a body-invariant input for the model. First, the RNNPB model is tested for robustness against noise and parametric noise in the original signals. It is demonstrated that the model is able to generalize, which means that the extended RNNPB model does not demand strict requirements on the features used for representing the state of the world. Furthermore, we demonstrated the present model in a toy world example, in which it is shown that observation of multiple goal-directed actions and generation of a goal directed action achieving the same goal both leads to equal activation of an additional input layer of the network. This layer contains the so called parametric bias nodes, which is an essential part of the RNNPB architecture. This exemplifies the goal related mirror properties of the model.
Faculteit der Sociale Wetenschappen