An approach to variable-order prediction via multiple distal dendrites of neurons

Abstract

In this paper, we proposed an extended version of binary code selection algorithm (BCSA) for the variable-order prediction by introducing multiple distal dendrites into BCSA. The proposed model of artificial neurons has a single proximal dendrite to receive the feed-forward inputs (sequences) from the world and multiple distal dendrites to receive the horizontal inputs from nearby neurons. During training, each distal dendrite is able to remember the states of neurons activated at different time and store the temporal correlations. After training, each distal dendrite independently recalls the temporal correlations contained in sequences and makes a local prediction. The variable-order prediction can be obtained by combining these local predictions made by multiple distal dendrites. Experiments show that the proposed method outperforms BCSA and other methods, such as back-propagation networks and radial basis function networks, especially while processing complex sequences.