Emotion illness is a kind of psychiatric disorder such as depression, anxiety disorders, and schizophrenia, which has an adverse effect on human society. In the simulation of the human emotion with device VDS level, the value range from -3 V to -25 V refers to the normal human emotions including happy, mild, and sad. When the VDS is decreased to an ultralow level (-1 V and -0.1 V), the device exhibits irregular responses to the same 10 consecutive light stimuli (0.3 s, 50 μW/cm2). With the decrease of the VDS, the baseline current of the device will also decrease (Figure \ref{857662}a). When the VDS is decreased to an ultralow level (-1 V and -0.1 V), the device exhibits irregular responses to the same 1- consecutive light stimuli (0.3 s, 50 μW/cm2) compared with the response of the device to light spikes under VDS = -3 V (Figure \ref{857662}b). The irregular response fluctuation of the device at VDS = -0.1 V and -1 V can be regarded as the synaptic unit in emotional illness. The damage of the depression in the human brain described in Figure \ref{857662}c inlcueds chaos, confusion, and irregular response to external stimuli. The external behaviors of the depression were successfully simulated by the current variations of our device to a certain degree.
The Emotion-SLP Simulation
The evaluate the device learning capability, the one-time potentiation and depression of our device was characterized. As shown in Figure S11, the light induced EPSC changes of our device can be effectively erased by the electric pulses, which indicates that our device can be reset to its initial value to carry out another round of new learning tasks without need to wait for a long time. For the further demonstration of the emotion-tunable learning capability of our device, we built a neural network based on single-layer perceptron (SLP) by using the extracted weight updating parameters from our neuromorphic device. We severe “emotion regulation” of SLP neural network learning capability by using the simulated neural network model to recognize modified national institute of MNIST handwritten digits after training. In Figure \ref{366439}a, the network would consist of 784 input neurons (the resolution of the digits image is 28×28) and 10 output neurons (the labels was set from 0 to 9) The input neurons and the output neurons are fully connected through 784×10 synapse (the values are regarded as synaptic weights). The input neuron would receive one signal converted from the gray level in one pixel of the training image (28×28). Then the input vector (Vi) functions through the weight values in the synaptic network (Wi, j). The calculation result was converted and transmitted to the output vector with the utilization of the sigmoid activation function. The difference between the image’s label and the output value would determine the direction of the synaptic-weight updating process via the backpropagation algorithm. Therefore, in one batch, the SLP network was trained through the input of 60000 image. Then the recognition rate (RR) of the trained SLP network for 10 numbers from “0” to “9” was tested with 10000 testing images. The curves and their fitting curves for the 100-weight update under various emotions (VDS, -5 V, -15 V) are shown in Figure \ref{366439}b. Before the simulation, the 100-times potentiation and depression at various VDS were tested (Figure \ref{366439}b). The corresponding curves were fitting using the following equations:
\(G_{n+1}=G_{n}+ΔG=G_{n}+α_{p}e^{-{β_{p}}\frac{G_{n}-G_{min}}{G_{max}-G_{min}}}\)
\(G_{n+1}=G_{n}-ΔG=G_{n}-α_{d}e^{-{β_{d}}\frac{G_{max}-G_{n}}{G_{max}-G_{min}}}\)
Where, Gn+1 and Gn stand for the conductance of the device when (n+1)th and nth pulses were applied, respectively. Gmax and Gmin represent the maximum and minimum conductance values. The parameters α and β indicate the step size of the condutance and nonlinearity, respectively. The SLP network was simulated based on the fitting parameters extracted from our device (Table S1). The device under higher VDS exhibits more linear potentiation and depression parameters than that under lower VDS. Therefore, the network under high VDS (positive emotion) exhibits a higher recognition rate (~75%) than that under low VDS (negative emotion, ~65%).