Lindong Wu

Lindong Wu

Auto-generated profile, based on publicly available information

Lindong Wu

School of Integrated Circuits, Peking University, Beijing, China

Research Focus & Metrics

Academic Career Length

6Years (2021-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
30 August 2021 in IEEE Transactions on Circuits and Systems II: Express Briefs
Verified
Journal Article
Closed access
YZYuhan Zhang
Yuhan Zhang
LWLindong Wu
Lindong Wu
WHWeihua He
Weihua He
ZZZiyang Zhang
Ziyang Zhang
CYChen Yang
Chen Yang
YYYing Yang
Ying Yang
This brief presents a novel spatio-temporal domain adaptation (DA) method that is unsupervised and event-driven for dynamic vision sensor (DVS) gesture recognition. This method realizes the transfer between the source domain and the target domain data in both the spatial and temporal dimension without the need of target domain data labels. Specifically, it consists of a deep spiking neural network (SNN)-based feature extractor, a label predictor and a domain discriminator. A time-space gradient reversal layer is responsible for building a spatio-temporal bridge between the domain discriminator and the feature extractor, which is essential to the alignment of the source domain spike features with the target ones and to achieve domain adaptation on both space and time dimensions. To demonstrate the effectiveness of our method, we adapted DVS hand gesture data from one temporal resolution to another and from original data to denoised data. Our method can provide up to 10.39% improvement in accuracy. Its accuracy improvement is more stable comparing with RNN-based and LSTM-based methods in this DA framework especially when the two domains have partial similarity in DVS data.

8 March 2022 in IEEE Transactions on Electron Devices
Verified
Journal Article
Closed access
LWLindong Wu
Lindong Wu
ZWZongwei Wang
Zongwei Wang
LBLin Bao
Lin Bao
ZYZhizhen Yu
Zhizhen Yu
QCQingyu Chen
Qingyu Chen
RHRu Huang
Ru Huang
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.

6 May 2024 in IEEE Transactions on Circuits and Systems II: Express Briefs
Verified
Journal Article
Closed access
YLYaotian Ling
Yaotian Ling
ZWZongwei Wang
Zongwei Wang
LWLindong Wu
Lindong Wu
YCYimao Cai
Yimao Cai
RHRu Huang
Ru Huang
Point clouds have important applications, such as computer vision, autonomous driving, and robotics. However, point cloud recognition task encounters challenges related to data input equivalence and computational overhead in conventional hardware. To address these issues, we propose a Hierarchical Synchronous Parallel Architecture (HSPA) for resistive-random-access-memory (RRAM) based computing-in-memory (CIM), which significantly enhances the computational parallelism of point cloud data with DeepSets network. The point cloud processing is experimentally implemented in RRAM-based HSPA CIM hardware fabricated using a commercial 40 nm CMOS technology. The results indicate that the HSPA achieves a remarkable energy efficiency of 9.70 fJ/op and a high framerate of 145.4 fps (@200MHz) while maintaining a software-comparable accuracy of 98.4%.

7 September 2021 in IEEE Journal of the Electron Devices Society
Verified
Journal Article
Open access
LWLindong Wu
Lindong Wu
ZYZhizhen Yu
Zhizhen Yu
YQYabo Qin
Yabo Qin
QCQingyu Chen
Qingyu Chen
YCYimao Cai
Yimao Cai
RHRu Huang
Ru Huang
The RRAM-based array is one of the most promising core functional primitives to accelerate the inference process of neural networks. However, the stress-induced disturbance can cause a significant accuracy drop during inference process where input vectors with different voltage levels are fed to the device. This kind of disturb can hardly be avoided by optimizing the fabrication process. Here, we investigate this phenomenon based on TaO x -based devices with different electrodes. The results indicate that the stress-induced disturb mainly appears in the intermediate resistance states when the voltage is applied on the device. The simulation result by COMSOL reveals the relationship between read disturb and electric field. Therefore, we propose a nonlinear weight quantification method to mitigate read disturb effect on inference accuracy by reducing the number of devices in intermediate resistance states. The simulation results based on the fully-connected networks for MNIST recognition indicate that stress disturb phenomenon can be well suppressed by nonlinear weight quantification compared with the conventional linear quantification method, which will advance the application of the RRAM-based accelerator.

8 April 2021
Verified
Conference Paper
Closed access
LWLindong Wu
Lindong Wu
ZWZongwei Wang
Zongwei Wang
ZYZhizhen Yu
Zhizhen Yu
YQYabo Qin
Yabo Qin
QCQingyu Chen
Qingyu Chen
RHRu Huang
Ru Huang
RRAM -based array has been utilized to accelerate inference process of neural network. However, the device resistance drift during inference will induce read disturb, which will cause significant accuracy drop and cannot be solved by process optimization. Here, we propose a nonlinear weight quantification method to mitigate read disturb effect on inference accuracy. The simulation results indicate that read disturb is well suppressed compared with traditional linear quantification method.

16 January 2024 in IEEE Transactions on Electron Devices
Verified
Journal Article
Closed access
YWYang Wang
Yang Wang
LWLindong Wu
Lindong Wu
CBChaoyi Ban
Chaoyi Ban
JYJian Yang
Jian Yang
XJXiangliang Jin
Xiangliang Jin
YWYuan Wang
Yuan Wang
Neuronal oscillation, as the fundamental component of information processing and transmission in the brain, plays a pivotal role in human cognition, learning, and memory. In this article, we introduce a novel silicon neuron device (SND) featuring light pulse modulation, aiming to emulate the oscillatory dynamics observed in biological neurons. The SND fabrication utilizes the 0.18 $\mu$ m standard Bipolar-CMOS-DMOS (BCD) process, with the incorporation of light source control enabling rapid firing of the neurons. The experimental findings demonstrate a strong correlation between the voltage oscillation characteristics of the device and the wavelength, intensity, and frequency of the light pulse. By adjusting the parameters of the pulses, diverse oscillation modes can be achieved. When the wavelength of the input light pulse signal is 480 nm and the intensity is 25 lx, the SND exhibits the capability to achieve stable excitation and quenching of the output pulse within the frequency range of 10 Hz–1 kHz. It demonstrates high sensitivity and strong responsiveness to the input signal. Furthermore, we exploit the strong dependence of the device on the excess bias voltage ( V $_{\text{EX}}$ ) to characterize the stochastic pulse firing of the SND in a light environment. Finally, the pulse firing characteristics of the device under different light source conditions were further quantified through letter recognition functional demonstrations, and it was applied to image preprocessing, effectively improving the accuracy of neural network for digital recognition.