MOTION LEARNING USING SPATIO-TEMPORAL NEURAL NETWORK
Keywords:Motion learning, reinforcement learning, reward-modulated spike-timing-dependent plasticity, spatio-temporal neural network
Motion trajectory prediction is one of the key areas in behaviour and surveillance studies. Many related successful applications have been reported in the literature. However, most of the studiesare based on sigmoidal neural networks in which some dynamic properties of the data are overlooked due to the absence of spatiotemporal encoding functionalities. Even though some sequential (motion) learning studies have been proposed using spatiotemporal neural networks, as in those sigmoidal neural networks, the approach used is mainly supervised learning. In such learning, it requires a target signal, in which this is not always available in some applications. For this study, motion learning using spatiotemporal neural network is proposed. The learning is based onreward-modulated spike-timing-dependent plasticity (STDP), whereby the learning weight adjustment provided by the standard STDP is modulated by the reinforcement. The implementationof reinforcement approach for motion trajectory can be regarded as a major contribution of this study. In this study, learning is implemented on a reward basis without the need for learningtargets. The algorithm has shown good potential in learning motion trajectory particularly in noisy and dynamic settings. Furthermore, the learning uses generic neural network architecture, whichmakes learning adaptable for many applications.