Kai-Ming Feng

Kai-Ming Feng

Auto-generated profile, based on publicly available information

Kai-Ming Feng

Department of Electrical Engineering, National Tsing Hua University, Hsinchu, Taiwan

Research Focus & Metrics

Academic Career Length

27Years (2000-Present)

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

Representative Publications

View All Publications
22 November 2018 in IEEE Photonics Technology Letters
Verified
Journal Article
Closed access
SSShuyi Shen
Shuyi Shen
JYJhih-Heng Yan
Jhih-Heng Yan
PPPeng-Chun Peng
Peng-Chun Peng
CHChin-Wei Hsu
Chin-Wei Hsu
QZQi Zhou
Qi Zhou
GCGee-Kung Chang
Gee-Kung Chang

16 April 2021 in Optics Letters
Verified
Journal Article
Closed access
PDF file
GYGovind Sharan Yadav
Govind Sharan Yadav
CCChun-Yen Chuang
Chun-Yen Chuang
KFKai-Ming Feng
Kai-Ming Feng
JCJyehong Chen
Jyehong Chen
YCYoung-Kai Chen
Young-Kai Chen
In this Letter, we propose and experimentally demonstrate a novel, to the best of our knowledge, sparse deep neural network-based nonlinear equalizer (SDNN-NLE). By identifying only the significant weight coefficients, our approach remarkably reduces the computational complexity, while still upholding the desired transmission accuracy. The insignificant weights are pruned in two phases: identifying the significance of each weight by pre-training the fully connected DNN-NLE with an adaptive L2-regularization and then pruning those insignificant ones away with a pre-defined sparsity. An experimental demonstration is conducted on a 112 Gbps PAM4 link over 40 km standard single-mode fiber with a 25 GHz externally modulated laser in O-band. Our experimental results illustrate that, for the 112 Gbps PAM4 signal at a received optical power of − 5 d B m over 40 km, the proposed SDNN-NLE exhibits promising solutions to effectively mitigate nonlinear distortions and outperforms a conventional fully connected Volterra equalizer (VE), conventional fully connected DNN-NLE, and sparse VE by providing 71%, 63%, and 41% complexity reduction, respectively, without degrading the system performance.

7 December 2020 in Optics Express
Verified
Journal Article
Closed access
GYGovind Sharan Yadav
Govind Sharan Yadav
CCChun-Yen Chuang
Chun-Yen Chuang
KFKai-Ming Feng
Kai-Ming Feng
JYJhih-Heng Yan
Jhih-Heng Yan
JCJyehong Chen
Jyehong Chen
YCYoung-Kai Chen
Young-Kai Chen
Volterra equalization (VE) presents substantial performance enhancement for high-speed optical signals but suffers from high computation complexity which limits its physical implementations. To address these limitations, we propose and experimentally demonstrate an elastic net regularization-based pruned Volterra equalization (ENPVE) to reduce the computation complexity while still maintain system performance. Our proposed scheme prunes redundant weight coefficients with a three-phase configuration. Firstly, we pre-train the VE with an adaptive EN-regularizer to identify significant weights. Next, we prune the insignificant weights away. Finally, we retrain the equalizer by fine-tuning the remaining weight coefficients. Our proposed ENPVE achieves superior performance with reduced computation complexity. Compared with conventional VE and L1 regularization-based Volterra equalizer (L1VE), our approach show a complexity reduction of 97.4% and 20.2%, respectively, for an O-band 80-Gbps PAM4 signal at a received optical power of −4 dBm after 40 km SMF transmission.

20 June 2022 in IEEE Access
Verified
Journal Article
Open access
PTPouya Torkaman
Pouya Torkaman
GYGovind Sharan Yadav
Govind Sharan Yadav
PWPo-Chuan Wang
Po-Chuan Wang
TLTung-Yu Lu
Tung-Yu Lu
XMXuan-Wei Miao
Xuan-Wei Miao
Shang-Hua YangShang-Hua Yang
Shang-Hua Yang
In this paper, a heterogeneous communication system capable of delivering 5G/sub-terahertz signal carriers over an arbitrary long fiber and separated transmission links is presented by employing direct detection, multiplexing techniques, and advanced digital signal processing. In this experiment, the 3.5 GHz and 28.5 GHz carrier frequencies, representing 5G links, deliver 4 Gb/s 16-QAM OFDM signals to separate user ends over a 1-meter wireless link distance. Later, the sub-terahertz wireless communications of 4 to 10 Gb/s QPSK and 8-QAM signals with varying carrier frequencies of 125-,175- and 225 GHz, over wireless distances (< 80 cm) are presented and evaluated. The results indicate that by increasing optical power from 12 dBm to 13 dBm, the bit error rate decreases two orders of magnitude. Eventually, with the assistance of artificial intelligence, a nonlinear equalizer (AI-NLE) prototype is introduced. The results indicate that the AI-NLE successfully decreases the number of errors in received data by one order of magnitude. The proposed heterogeneous system is compatible with radio-over-fiber technology, cost-effective, and easy to deploy, making it a promising candidate for indoor terahertz communication.

24 June 2024 in IEEE Communications Letters
Verified
Journal Article
Open access
PTPouya Torkaman
Pouya Torkaman
SLSeyed Mostafa Latifi
Seyed Mostafa Latifi
KFKai-Ming Feng
Kai-Ming Feng
Shang-Hua YangShang-Hua Yang
Shang-Hua Yang
Linear and nonlinear impairments restrict the transmission performance of high-speed terahertz (THz) communication systems. To improve transmission performance, we propose a two-stage nonlinear equalizer (NLE). In the first stage, a memory-controlled long short-term memory (LSTM) neural network learns channel nonlinearity and compensates for it through nonlinear waveform regression. In the second stage, a low-complexity deep random forest (RF) network identifies nonlinear boundaries among individual QAM symbols and adjusts the standard hard decision thresholds of the QAM demodulator to align with the distribution of received symbols. This study experimentally validates the proposed two-stage NLE on a dual-channel THz-over-fiber transmission system using an intensity modulation and direct detection (IM/DD) scheme, achieving a successful 20 Gbps line rate up to a 4.5-meter wireless link at both 125/300 GHz frequency bands. The proposed scheme outperforms a Volterra nonlinear equalizer in all tested scenarios, surpassing a linear equalizer (LE) by reducing the bit error rate (BER) from 2.47 × 10 -3 to 2.61 × 10 -4 in the 300 GHz link and from 3.42 × 10 -3 to 5.64 × 10 -4 in the 125 GHz link.

20 December 2022 in Journal of Lightwave Technology
Verified
Journal Article
Closed access
GYGovind Sharan Yadav
Govind Sharan Yadav
CCChun-Yen Chuang
Chun-Yen Chuang
KFKai-Ming Feng
Kai-Ming Feng
JCJyehong Chen
Jyehong Chen
YCYoung-Kai Chen
Young-Kai Chen
In this paper, we propose and experimentally validate a novel sparsity learning deep neural network nonlinear equalization method (SLDNN-NLM) for a 112-Gbps PAM4 signal using a 25-GHz externally modulated laser (EML) to transmit over 40-km standard single-mode fiber (SSMF) in O-band. Our proposed SLDNN-NLM can reduce computation complexity while still maintain the system performance. This method combines the L2-regularization term into the loss function and adapts the deep neural network (DNN) model by distinguishing the significance of each connection. We next prune away insignificant weight connections using the pruning technique without effectively degrading the system performance. In our experimental validation, we compare the system performance and computation complexity of our approach with conventional deep neural network based nonlinear equalizer (CDNN-NLE), L2-regularized Volterra nonlinear equalizer (L2VNLE), sparse CDNN-NLE, and sparse L2VNLE. The experimental findings imply that the proposed method has great potential to deal with nonlinear distortions arisen from the optical transmission systems. Our results show that the proposed SLDNN-NLM with L2-regularization achieves better BER performance than CDNN-NLE at the same complexity level. Moreover, when comparing with L2VNLE, the computation complexity is reduced by 42% at each received optical power (ROP) after 40-km SSMF transmission. Compared with L2VNLE, CDNN-NLE, sparse L2VNLE, and sparse CDNN-NLE, the overall complexity of the proposed SLDNN-NLM after pruning can be reduced by 86%, 75%, 70%, and 61% for BTB at an ROP of −3dBm, and 79%, 63%, 58%, and 57% for 40-km transmission at an ROP of −5 dBm, respectively, without degrading the system performance.