A Study on a Cortical Learning Algorithm Dynamically Adjusting Columns and Cells

Abstract

The cortical learning algorithm is one of the time-series data prediction algorithms based on the human neocortex. The cortical learning algorithm uses multiple columns to represent an input data value at a time step, and each column has multiple cells to represent a time-series context of the input data. In the conventional cortical learning algorithm, the numbers of columns and cells are user-defined parameters. Their appropriate numbers depend on the input data, and the input data are unknown before the learning. To avoid the parameter settings of the number of columns and cells, in this work, we propose a self-structured cortical learning algorithm dynamically adjusting columns and cells according to the input data. Experimental results using test time-series input involving a sine, a combined sine, and a logistic mapping data show that not only the proposed self-structured algorithm can dynamically adjust columns and cells depending on the input data but also the prediction accuracy is also higher than the conventional LSTM and CLA.