Fingerprint-Based Training and Evaluation on Hierarchical Temporal Memory

In spiking neural networks and biologically realistic simulation neural networks, Hierarchical Temporal Memory (HTM) has been proposed, yet it encounters challenges related to storage capacity. This study introduces an enhancement to the training process, which modifies the dendrite weights based on the evaluation of hash performance and the achievement of learning and recognition. The training process is optimized through the bionic hashing mechanism of BioHash, utilizing the Energy function to assess the payload of HTM modules. This assessment facilitates the distribution of storage requirements among different modules. On the CIFAR-10 dataset, we conducted classification tests using the proposed Hash-HTM with different fingerprint size of hashes, demonstrating a significant improvement at long hash cases in accuracy to illustrate the effectiveness of the storage enhancement.