Selective Dynamic Principal Component Analysis Using Recurrent Neural Networks

In the last decades, considerable attention has been focused on development of bio-inspired systems. This paper employs the principals of information processing in the Basal Ganglia (BG) to develop a new method for selectively extracting dynamic principal components (DPCs) of multidimensional datasets. The DPCs are extracted by are current structure of auto-associative neural network and selectivity is achieved by means of a reinforcement-like signal which modifies the desired outputs and the learning coefficient of the network. Performance of the model is evaluated through two experiments; at first, the DPCs of a stock price database are extracted and then, speech compression capability of the method is checked which illustrates the efficiency of the proposed approach.