10 June 2019 in IEEE Electron Device Letters
We propose an electronic synapse based on an Al: HfO2 metal-ferroelectric metal (MFM) capacitor with multi-level characteristics. The device demonstrates excellent multilevel behavior, linearity, weight update symmetry, and low static power consumption. Further, an artificial neural network based on this capacitor is built and evaluated through SPICE and MATLAB simulations. The performance indicates that higher than 78% accuracy and nJ/ms dynamic speed-energy efficiency in recognizing the standard MNIST handwritten digits are achieved, showing the capability of high-density integration and low energy consumption of capacitive artificial neural networks.
19 August 2020 in IEEE Transactions on Electron Devices
In a digital-analog mixed neuromorphic system, various complex peripheral circuits may offset the integration and energy efficiency advantages of the dense crossbar. To simplify the peripheral circuits, a self-activation neural network (SANN) is proposed based on a passive crossbar array formed by the rectified memristive (ReMem) cell. The ReMem cell is a dual-mode resistive switching device with a VO₂/TaOₓ bilayer structure. It shows a hybrid switching behavior including volatile threshold switching and multilevel nonvolatile resistive switching. This unique feature enables not only simultaneously self-selection and distributed activation, but also weight storage. The concept of SANN is theoretically examined, and a neural network is constructed and trained with SANN based on the proposed devices to perform recognition on handwritten digits in the MNIST database. Compared with neuromorphic computing systems using the CMOS-based activation module and additional selective element, the results show comparable recognition accuracy with reduced circuitry complexity. The proposed SANN can be a promising alternative to realize neural network computing systems with simplified peripheral circuits.
10 April 2020 in IEEE Electron Device Letters
In this letter, a new reliability phenomenon, named early-stage resistance fluctuation (ERF), in analog resistive random-access memory (RRAM) devices and its impact on the neuromorphic computing are investigated. ERF is found to be a non-negligible random fluctuation behavior of RRAM resistance within ~10,000s after verification due to stochastic migration of vacancies. As a result, ERF can induce weight noise in RRAM-based neural network even with write-verify operation and cause a significant (~42%) drop of recognition accuracy on the MNIST dataset recognition tasks. An approach incorporating automatic early-stop and dropout during training is proposed to reduce the impacts of ERF. Results show that the recognition accuracy with ERF can be improved from 58% to 96% after adopting the proposed optimization method.
1 January 2021 in IEEE Electron Device Letters
Compressed sensing with a tunable random sampling strategy has great potential in reducing the energy/time consumption during the constant acquisition of external information, thereby it is considered one of the most promising sensing strategies in edge-terminal. In this paper, a novel tunable stochastic oscillator (TSO) based on the VO2/TaOx structure was proposed and employed as the core controlling device for compressed sensing. The TSO is jointly governed by the metal-insulator-transition (IMT) based threshold switching of the VO2 layer and the stochastic resistive switching by the TaOx layer. The tunable memristor/resistor in series can modulate the oscillation frequency of the TSO. The orthogonal matching pursuit (OMP) algorithm is implemented to reconstruct the compressed sparse signals, demonstrating a good performance of TSO in compressed sensing.
8 March 2022 in IEEE Transactions on Electron Devices
Neurons are basic elements of the human brain to transmit and process various kinds of information. Besides the synaptic plasticity, the neuronal intrinsic plasticity (NIP) plays a vital role in improving and stabilizing the neural circuit for learning and adaptation to environment. In this article, we emulate the NIP behavior by combining hybrid VO 2 $/$ TaO x device with a tunable load resistor (TLR). By adjusting the TLR and stimulus intensity, the oscillatory threshold of artificial neurons can be adaptively modulated. Moreover, inspired by the nervous system, we implement the artificial NIP as an intercept-adjustable rectified linear unit (IA-ReLU) in the conversion-based spiking neural network (SNN). The results indicate that the NIP can mitigate the stuck-at-fault (SAF) effect on a memristor array and improve the recognition accuracy by up to 4.4% compared with the traditional ReLU function. These results demonstrate that the intrinsic plasticity of neurons can play a vital role in spiking neuromorphic computing systems.
15 April 2022 in IEEE Transactions on Electron Devices
Pulse-to-pulse variability (PPV) is a vital nonideal effect in analog resistive memory, particularly for online training in emerging computing paradigms. The reported PPVs usually show two seemingly contradictory properties: large PPV amplitude between neighboring write pulses but statistically stable switching curves. In this article, a new insight is proposed to unambiguously decouple PPV into two independent components: variability and curve fluctuation. Furthermore, a phenomenological modeling and parameter extraction method is proposed to capture this characteristic. The proposed model can facilitate the exploration of PPV impact on neural network accuracy for online training.