Defect-Resilient Memristor Crossbar of Hierarchical Temporal Memory (HTM) Spatial Pooling for Near-IoT-Sensor Cognitive Computing

摘要

A huge amount of IoT-sensor data may threaten a capacity limitation of cloud computing in near future. To mitigate computation burden of cloud computing due to the massive amount of sensed data, a memristor crossbar for Hierarchical Temporal Memory (HTM) spatial-pooling can be considered for energy-efficient near-IoT-sensor computing. By doing so, the amount of data sent to the cloud server can be reduced significantly from 784 gray pixels to 400 Sparse Distributed Representation (SDR) bits or 256 SDR bits for processing MNIST hand-written digits. The loss of recognition rate is as little as 2.48% and 1.93% for the 256-column and 400-column memristor crossbars of spatial-pooling, respectively. In addition, we have tested and verified the defect-resilience scheme of spatial-pooling memristor crossbar in this paper, because memristor crossbars near IoT sensor are vulnerable to memristor defects such as stuck-at-faults, resistance variations, etc.