Nonmasking-based reservoir computing with a single dynamic memristor for image recognition

Reservoir computing has been widely used in temporal information processing, and the presentation of time-delayed reservoir computing systems effectively reduces the difficulty of the physical implementation of reservoir computing. However, challenges of complex structure and difficult multi-parameter optimization still persist. To address these issues, this study simplifies the structure of time-delayed reservoir computing system by removing the masking procedure and feedback loop and proposes a nonmasking-based reservoir computing system using only a single dynamic memristor for image recognition tasks. The histogram of oriented gradient (HOG) feature of the input image is linearly mapped into a voltage sequence and directly injected into a dynamic memristor. The nonlinear mapping of the input signal is performed by utilizing the physical computing resources of the dynamic memristor, so as to construct a reservoir computing system without masking procedure and feedback loop. The proposed reservoir computing system achieves effective image recognition only by properly adjusting the range of the mapping voltage. The recognition accuracies on the image recognition tasks of MNIST and Fashion-MNIST datasets are \(98.44\%\) $98.44\%$ and \(90.19\%\) $90.19\%$ , respectively, surpassing the same type dynamic memristor-based parallel reservoir computing system and laser-based reservoir computing system. Moreover, the recognition accuracy on the MNIST dataset is only slightly reduced by \(0.14\%\) $0.14\%$ than that of the classical reservoir computing system with 1200 physical nodes. In comparison to the proposed reservoir computing system with masking procedure, the training time of the proposed nonmasking-based reservoir computing system on the MNIST and Fashion-MNIST datasets is reduced by about \(46.1\%\) $46.1\%$ and \(45.33\%\) $45.33\%$ , respectively, while the corresponding recognition accuracy is only slightly decreased by \(0.5\%\) $0.5\%$ and 0.35 \(\%\) $\%$ , respectively. In addition, the experimental results on CIFAR-10 and Cropped SVHN datasets further verify the feasibility of the proposed reservoir computing system in complex image recognition tasks.