This very comprehensive technical account presents a thoroughly worked out “architecture for the spatial pooler (SP)” of the hierarchical temporal memory (HTM) algorithm. This algorithm is designed to produce invariant representations of spatial and spatio-temporal inputs.

The proposed research defines different features that act on HTM synapses, such as memristors, synthetic synapses, and a crossbar architecture for the SP to optimize algorithmic performance during the training phase. Furthermore, to strengthen the robustness of the SP, plasticity mechanisms are introduced: neurogenesis and homeostatic intrinsic plasticity.

This well-argued paper includes a detailed overview of HTM and a step-by-step description of the algorithm and its components. It first presents the adopted design methodology for improving the algorithm, with a lot of technical insight and clearly explained formulas. Next, it covers system design and implementation in a way that is reproducible. Finally, it outlines the experimental setup in equally great detail, including a discussion of the evaluation metrics. The experiments are carried out in the image recognition domain over well-established datasets (the Modified National Institute of Standards and Technology (MNIST) database and the Yale Face database), evaluating noise robustness and power consumption, for example, to prove the algorithm’s suitability for mobile devices. The evaluation results section describes the high scores without comparing the proposed approach to other approaches.

The paper is clear and well done, with a long list of references and a well-described (reproducible) algorithm. It is a very good read for graduate students, scholars, and engineers interested in artificial intelligence and optimal methods for machine intelligence.