2.4. Full-hardware neuromorphic vision system based on reconfigurable ion-modulated memtransistors
Since the directional movement of ions in the electrolyte needs to surmount potential energy barrier, and the additional silicon oxide layer also causes an inevitable voltage drop, the short-term response of the device has a nonlinear relationship with the amplitude of external stimuli. To investigate the nonlinear response, we applied a series of identical pulses to the device, ranging in amplitudes from 0.2 V to 3.6 V. The corresponding increases in drain current were shown in Figure 5a. The drain current at the end of the last stimuli pulse and the first pulse stimuli were summarized in Figure 5b and 5c, respectively. To take the relation between the drain current responses and stimuli amplitude into systematic computation, a softplus-like function ( y=aln(1+ebx) ) was adopted to fit the experimental results.
In artificial vision system, images captured by image sensor are often distorted by various noises, such as electrical noise, mechanical noise, channel noise and other noises, during generation and transmission. To suppress noise, improve image quality, and facilitate higher-level processing, image denoising is performed using the short-term dynamics of the ion-modulated memtransistors, as illustrated in the inset of Figure 5b. During inference, the MVM is often performed by applying a short pulse on the bit line. We use the channel current change characteristic after the single-pulse stimulus in the operation of neuron activation in the inference, which is shown in the inset of Figure 5c. The schematic of the basic neural network architecture for the neuromorphic vision system is demonstrated in Figure 5d, mainly including filtering units for denoising, synapses for MVM and hardware softplus neurons for nonlinear activation. Then the artificial neuromorphic hardware systems for visual information processing were proposed based on the ion-modulated memtransistors, as shown in Figure 5e. After stimulating by the encoded electrical pulses in the filtering units, the drain currents were transferred into voltage pulses in a linear mapping relation. Then the converted voltage pulses were fed into the Computing-in-Memory array to perform MVM. The basic cell consists of one transistor and one ion-modulated memtransistor (1T1M), in which the transistor is responsible for selective programming and retention enhancing. Finally, the cumulative current after the MVM is converted to voltage pulses and then applied on the gate to utilize the softplus-like response to achieve the nonlinear activation.
As shown in Figure 6 a, we simulated multilayer perception (MLP, inset in Figure 6a) for the evaluation of the network-level performance using the ion-modulated memtransistor for softplus neurons. The simulation details can be found in the Experimental Section, there was almost no difference in the testing accuracy between the standard software softplus neuron and the hardware softplus neuron. Unlike weight updating, there is supposed to be no accumulation in the device state for the application of neuron function. To avoid the transition between short-term memory and long-term memory, it is necessary to impose a constraint on the amplitude of gate pulses. We define the viable upper limit amplitude of the gate pulses as cutoff voltage, with minimizing the cutoff voltage, the accumulative effect can be overcome and programming energy can be saved. As shown in Figure 6b and 6c, there is no significant difference in the network performance with cutoff voltage in the interval between 3.0 V and 3.5 V. However, a noticeable degradation took place when the cutoff voltage reached 2.9 V. Moreover, after reducing the cutoff voltage below 2.8 V, there is no classification ability for the neural network, with the accuracy all about 10%. After investigating the adequate upper limit of gate pulses, we set the cutoff voltage as 3 V. To characterize the endurance of the devices, a train of pulses of 3 V was applied on the device gate. Figure 6d shows no sign of ON/OFF ratio degradation up to 2000 cycles, implying that there is no need to reset the device to the initial state with the help of peripheral circuits. The spontaneous decaying characteristic can ensure repeated activations in the inference. As for the noisy nature of the diffusion of the random ions, we explore the influence on the neural network performance under the noise of the hardware softplus neuron. It can be seen in Figure 6e that the neural network can tolerate the considerable noise level of the hardware softplus neuron, implying the robustness for the ion-modulated memtransistor configuring as the neuron function, the hardware softplus functions with Gaussian noise of different standard deviation were also compared in Figure S7a, Supporting Information, and the collection of testing accuracies for the network concerning the neuron function was demonstrated in Figure S7b, Supporting Information.
Following the discussions about hardware softplus neuron implementation, we choose the cutoff voltage of 3 V and noise level of 10% for the subsequent investigation of the filtering function of ion-modulated memtransistors. Firstly, we compared the same images in three different states: 1) Original; 2) With 10% Gaussian noise; 3) After softplus-like function filtering; and the results are shown in Figure 6f. Compared with the noisy images, after filtering, background noises could be suppressed and critical image information got enhanced. Although there was an overall reduction in the specific pixel value, the shape could still be distinguished by the enhanced contrast with the background noises. As shown in Figure 6g, after filtering by the device nonlinearity, the testing accuracy got a significant increase from 11.66% to 77.96%, and the corresponding confusion matrix in Figure 6h demonstrated improved classification accuracy by filtering the image noises. One of the determining factors of the specific filtering function is the mapping gate voltage range. As illustrated in Figure S8, Supporting Information, there are some differences among the filtering functions of different starting mapping gate voltages. As shown in Figure 6i, the suitable range of the starting mapping gate voltages was located between 1.5 V and 1.8 V. In the following investigations, the filtering function started mapping from 1.5 V and ended at 3 V. Moreover, one of the typical testing images with different levels of Gaussian noises was demonstrated in Figure 6j, and the final testing accuracies among the original images, noisy images and filtered images were compared in Figure 6k, implying the neuromorphic vision systems can process images with considerable noises which is even hard for the human being to recognize. Finally, we also discussed the impact of different drain biases on device performances. As shown in Figure S9, Supporting Information, with the increase of drain biases, the decaying speed also got boosted, but there were no notable differences in the decaying characteristics when the biases were beyond 0.3 V. Although increased decaying speed can help reduce the delay in the inference, as shown in Figure S10, Supporting Information, the drain current also got increased with the higher drain bias, which will cause extra energy costs for computing. Therefore, final drain bias was set at 0.2 V for tackling the dilemma.