Predictive coding for dynamic visual processing

Abstract

The current paper proposes a novel predictive coding type neural network model, referred to as the predictive multiple spatio-temporal scales RNN (P-MSTRNN). P-MSTRNN learns to predict visually perceived whole-body cyclic movement patterns of human by exploiting multiscale spatio-temporal constraints which are imposed on the network dynamics by using different size of receptive field as well as different time constant value at each layer. The network after learning becomes able to proactively imitate movement patterns of targets by inferring or recognizing their intention by means of regression of the prediction error. We examined how the performance of the model in pattern generation as well as predictive imitation varies depending on stage of learning. The results showed that the network can develop functional hierarchy by developing different class of dynamic structure at each layer. Examination of the developmental processes clarified that number of limit cycle attractor of embedding target movement patterns increases as the learning proceeds. It was, however, shown that the tasks of pattern generation and predictive imitation can be performed successfully by using even the transient dynamics developed in the early stage of learning. From this result the paper concludes that exploitation of transient dynamics is beneficial for achieving successful task performance by conducting only limited period of learning.