Fig. 4: Curvature discrimination.

a Schematic flow of the curvature discrimination task implemented on the all-ferroelectric RC system. Each input curve is chopped into 3 sections and then converted to 3 3-timeframe pulse trains, which are applied to a 3-volatile-FD reservoir. The reservoir states are subsequently fed to a nonvolatile FD-based readout network to obtain the final output. Note that only 1 column of nonvolatile FDs in the network is used in this task because it is a simple binary classification task. b Photo of the experimentally constructed all-ferroelectric RC system. c Experimentally measured reservoir states after presenting 10 typical curves from the test set to the 3-volatile-FD reservoir. d Output currents generated by the output neuron in the readout network during the presentations of the 10 typical curves. The error bars are obtained from 4 independent experiments. e Neuronal outputs after feeding the output currents in (d) to the sigmoid activation function. f Comparison of the accuracies on the test set (36 curves) between the all-ferroelectric RC system with a volatile FD-based reservoir and two control RC systems with linear resistor- and sigmoid-based reservoirs.