Motor experience alters action perception through predictive learning of sensorimotor information

Recent studies have revealed that infants' goal-directed action execution strongly alters their perception of similar actions performed by other individuals. Such an ability to recognize correspondences between self-experience and others' actions may be crucial for the development of higher cognitive social skills. However, there is not yet a computational model or constructive explanation accounting for the role of action generation in the perception of others' actions. We hypothesize that the sensory and motor information are integrated at a neural level through a predictive learning process. Thus, the experience of motor actions alters the representation of the sensorimotor integration, which causes changes in the perception of others' actions. To test this hypothesis, we built a computational model that integrates visual and motor (hereafter, visuomotor) information using a Recurrent Neural Network (RNN) which is capable of learning temporal sequences of data. We modeled the visual attention of the system based on a prediction error calculated as the difference between the predicted sensory values and the actual sensory values, which maximizes the attention toward not too predictable and not too unpredictable sensory information. We performed a series of experiments with a simulated humanoid robot. The experiments showed that the motor activation during self-generated actions biased the robot's perception of others' actions. These results highlight the important role of modalities integration in humans, which accounts for a biased perception of our environment based on a restricted repertoire of own experienced actions.