Variation-Aware Non-linear Mapping for Honey-Memristor Based Neuromorphic System

Natural organic memristors are promising synapse candidates for neuromorphic systems due to their significant benefits, such as environmental sustainability, low production and disposal cost, non-volatile storage capability, and bio/CMOS-compatibility. In this paper, experimental evaluations of of a neuromorphic system based on honey-memristors are reported. First, honey-memristors are manufactured and tested based on our in-house technology, and then the non-linear characteristics inherent to honey-memristors, which causes inaccurate weight updates and reduces the inference accuracy, are explored. A non-linear mapping (NMP) method is applied to mitigate the effects of non-linearity in honey-memristor, under different scenarios including multiple cycle-to-cycle variations and Analog-to-Digital Converter (ADC) quantization. Experimental results indicate that, using 4- and 5-bit ADC, the inference accuracy of the neuromorphic system increases up to 13.3% without cycle-to-cycle variations and up to 18.6% with cycle-to-cycle variations through the NMP method. Finally, a comparison with state-of-the-art is presented to show merits of our proposed system which serves as a strong foundation to encourage further research into natural organic memristors for neuromorphic systems.