Column-based Decoder of Internal Prediction Representation in Cortical Learning Algorithms

Abstract

This work proposes a decoder of internal prediction representation in the cortical learning algorithm (CLA). CLA is a time-series data prediction algorithm. CLA receives an input bit-string at each time-step and encodes it to an internal data representation based on the relationship between the input data bits and columns, which are components of the CLA predictor. To decode the internal predictive presentation to the same format of the input bit-string, the conventional CLA employs the sparse distributed representations classifier (SDRC), which uses the relationship between the input data bits and cells, which are other components of the CLA predictor. To build the relationship between the input data bits and the cells, the conventional SDRC additionally needs a time-consuming learning process and a learning parameter, and the prediction accuracy cannot be expected during its learning process. In order to overcome these problems of the SDRC decoder in the conventional CLA and improve the prediction accuracy, the proposed method decodes the prediction representation inside CLA by utilizing the relationship between the input data bits and the columns built during the encoding process. The proposed method avoids the additional learning process to build the relationship between the input data bits and the cells. Also, The proposed method avoids the setting of any learning parameter needed in the conventional SDRC decoder. Experimental results using artificial benchmark and real-world electricity load time-series data show the proposed column-based decoder contributes to rapidly decreasing the prediction error and achieving smaller prediction error than the conventional SDRC decoder.