Feature extraction using neocognitron learning in Hierarchical Temporary Memory

Hierarchical Temporal Memory (HTM) serves as a practical implementation of the memory prediction theory. In order to obtain the optimum accuracy in pattern recognition, it is crucial to apply an appropriate learning algorithm for the feature extraction step of the HTM. This study proposes the use of neocognitron learning in extracting features of the pattern for image recognition. The integration of neocognitron into HTM addresses both the scale and time issues of the HTM. As for evaluation, a comparison is made against the original HTM and principal component analysis (PCA). The results show that more features are extracted as a function of input patterns than the original HTM and PCA.