Pattern recognition using hierarchical concatenation

The book On Intelligence written by Jeff Hawkins proposes a theory of how intelligent behaviour is generated in the brain´s neocortex. The underlying mechanism is based on a layered structure called the Hierarchical Temporal Memory (HTM). Researchers have been reporting software and hardware implementation of the HTM, with different levels of success. This paper presents a pattern recognition system based on Hawkins´ concept of simple concatenation, where decisions are made based on the processing of information in a hierarchical array of cells. A two-layer network recognizing 32×32 pixel patterns was used as a proof of concept. In this implementation patterns on the upper layer are activated by concatenating patterns presented on the lower layer. The input pattern is analysed in the lower layer by moving a coincidence array. Outputs from the lower layer are fed to the higher layer where they are analysed with a bigger coincidence array. The results in this paper were obtained using a set of seven patterns. One of the patterns was used as a sample or learnt pattern, and the rest of the patterns were used for testing. The efficacy of the network was measured using the percentage of similarity between testing patterns and the learnt pattern. Two different ways of moving the coincidence array were explored. The first method moves the coincidence array by its size, and the second moves the coincidence array by one pixel at a time. For validation, results were also compared to the existing k-nn algorithm and human classification. Moving the coincidence array by one pixel produced a percentage of similarity closer to the results of k-nn, while moving the coincidence array by its size produced a classification closer to the human perception.