A model for biological motion detection based on motor prediction in the dorsal premotor area

Recent findings regarding dorsal premotor area (PMd) activation during observation of smooth biological movements suggest that this motor-related area detects biological motions. We hypothesize that a neural network in the PMd acquires an invariance of self-induced motor commands for smooth movements and interprets the observed biological motions as ones satisfying the invariance in self-movements. To verify our hypothesis, we developed a recurrent neural network (RNN) to be trained with smooth motor movements, and examined how the RNN acquires biological invariance. The results showed that predictive learning of the RNN contributed to invariance acquisition, which enabled it to detect biological motions. Our findings agree with the fact that the PMd originally functions as a motor predictor. Moreover, this RNN could judge the ankle and wrist trajectories of a walking human as biological regardless of the subject's sex and emotional state.