Fig. 5: MPBTFT-based ARC system for handwritten digit recognition.

a Schematic of the ARC system utilizing volatile DG MPBTFTs and nonvolatile FeTFTs. In the system, input pulses are applied to the DG MPBTFT-based physical reservoirs. The generated reservoir states are fed into the readout network, which is comprised of FeTFT-based synaptic devices. Subsequently, the output currents from the readout network are transmitted to the DG MPBTFT-based LIF neurons, which produce the final outputs of the ARC system. b Preprocessing step of the MNIST dataset. Input images are cropped and encoded into input pulse trains. The image divided into five distinct sections along the row direction, each comprising 4 pixels, is shown as an example. Each of these encoded input pulse trains is then applied to the DG MPBTFT-based physical reservoirs. The obtained reservoir states from both the DG MPBTFT- and linear resistor-based physical reservoirs are depicted on the right side. c Classification accuracy for the MNIST dataset utilizing several different device-based physical reservoirs. Each result signifies the average accuracy derived from performing the task five times, ensuring a robust performance evaluation by mitigating the influence of outliers and variability. The DG MPBTFT-based physical reservoirs achieve the highest accuracy, particularly when utilizing the TG. d Average accuracies achieved for various physical reservoirs. e Confusion matrix for the MNIST digit recognition task. The MPBTFT-based ARC system accurately distinguishes ten types of handwritten digits.