Neural Network Equalizer Research Articles

Interference Estimation Using a Recurrent Neural Network Equalizer for Holographic Data Storage Systems

Holographic data storage (HDS) utilizes the unique properties of light for writing and reading two-dimensional (2D) data from holographic media, providing significantly higher densities and faster data transfer rates than traditional storage media for short-term dependencies. With its ability to store terabytes of data in a single crystal, HDS has garnered attention as a promising candidate for next-generation storage technologies. However, the 2D interference caused by hologram dispersion during the reading process poses a significant obstacle to achieving reliable and efficient HDS systems. This study proposes a method for enhancing the accuracy of estimating the 2D intersymbol interference (ISI) using a recurrent neural network (RNN) equalizer for HDS systems. The proposed method leverages the ability of RNNs to model complex and temporal dependencies in data and more accurately estimate the interference caused by ISI and interchannel interference (ICI) in HDS systems. In addition, to recreate the relationship between the samples in the training process, RNN is applied to fields such as computer vision, natural language process, speech recognition, and so on. We evaluated the performance of our proposed method on a simulation model of HDS system and compared it with the previous studies. In the simulations, the proposed method outperformed the previous schemes in terms of bit error rate, indicating its potential for improving the reliability and efficiency of HDS systems.

90-m/660-Mbps Underwater Wireless Optical Communication Enabled by Interleaved Single-Carrier FDM Scheme Combined With Sparse Weight-Initiated DNN Equalizer

In this paper, we first utilize the interleaved single-carrier frequency division multiplexing (I-SC-FDM) scheme with a novel sparse weight-initiated deep neural network (SWI-DNN) equalizer for underwater optical communication (UWOC) systems. In addition to the superior characteristics of I-SC-FDM over the original OFDM and localized SC-FDM (L-SC-FDM), such as lower peak-to-average power ratio (PAPR) and lower computational complexity, the implementation of a special sparse weight-initiated (SWI) structure can significantly reduce the necessary training epochs up to 10.3%, which enables the proposed SWI-DNN equalizer to outperform the traditional random weight-initiated DNN equalizer. Besides, further pruning operation can dramatically enhance the final sparsity of the SWI-DNN equalizer to 96.88% at the expense of an inappreciable performance penalty, thereby effectively saving occupied computing resources. Following this, a data rate of up to 660 Mbps over a 90-m underwater transmission can be achieved in a standard 50-m swimming pool, with a relatively low received optical power of -31.12 dBm (at the BER of 3.8 × 10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup> ). Such a data rate is 17.9% higher than that obtained with a common hybrid time-frequency domain (TFD) equalizer. Moreover, the computational complexity of the SWI-DNN equalizer after pruning operation is only 11.83% of that of the hybrid TFD equalizer and can still support a 625-Mbps transmission. This is the first time to employ I-SC-FDM combined with an SWI-DNN equalizer for long-distance high-speed UWOC transmission, and it can be highly beneficial to the cost-sensitive and power-sensitive systems in future deployment.

Spectrally stitched WDM nonlinear frequency division multiplexed transmission system

We present the concept and design of a 100 GHz PIC-based nonlinear frequency division multiplexed transmission system in which four b-modulated spectrally overlapping channels are spectrally stitched without frequency guard bands to obtain a continuous spectrum. Neural network-based equalizers are used at the receiver for improved performance. Simulations show that the proposed system alleviates the problem of nonlinear interaction between neighboring channels, that would, otherwise, be out of reach even for neural network equalization to solve on its own. The system achieves a 960 km transmission distance with a raw data rate of 333 Gbps per polarization and a BER well below the SD-FEC limit for every subcarrier.

Reducing Computational Complexity of Neural Networks in Optical Channel Equalization: From Concepts to Implementation

In this paper, a new methodology is proposed that allows for the low-complexity development of neural network (NN) based equalizers for the mitigation of impairments in high-speed coherent optical transmission systems. In this work, we provide a comprehensive description and comparison of various deep model compression approaches that have been applied to feed-forward and recurrent NN designs. Additionally, we evaluate the influence these strategies have on the performance of each NN equalizer. Quantization, weight clustering, pruning, and other cutting-edge strategies for model compression are taken into consideration. In this work, we propose and evaluate a Bayesian optimization-assisted compression, in which the hyperparameters of the compression are chosen to simultaneously reduce complexity and improve performance. Next, this paper presents four distinct metrics (RMpS, BoP, NABS, and NLGs) that are discussed here that can be used to evaluate the amount of computing complexity required by various compression algorithms. These measurements can serve as a benchmark for evaluating the relative effectiveness of various NN equalizers when compression approaches are used. In conclusion, the trade-off between the complexity of each compression approach and its performance is evaluated by utilizing both simulated and experimental data in order to complete the analysis. By utilizing optimal compression approaches, we show that it is possible to design an NN-based equalizer that is simpler to implement and has better performance than the conventional digital back-propagation (DBP) equalizer with only one step per span. This is accomplished by reducing the number of multipliers used in the NN equalizer after applying the weighted clustering and pruning algorithms. Furthermore, we demonstrate that an equalizer based on NN can also achieve superior performance while still maintaining the same degree of complexity as the full electronic chromatic dispersion compensation block. We conclude our analysis by highlighting open questions and existing challenges, as well as possible future research directions.

400 Gbit/s 4 mode transmission for IM/DD OAM mode division multiplexing optical fiber communication with a few-shot learning-based AffinityNet nonlinear equalizer.

Nonlinear impairment in a high-speed orbital angular momentum (OAM) mode-division multiplexing (MDM) optical fiber communication system presents high complexity and strong stochasticity due to the massive optoelectronic devices. In this paper, we propose an Affinity Network (AffinityNet) nonlinear equalizer for an OAM-MDM intensity-modulation direct-detection (IM/DD) transmission with four OAM modes. The labeled training and testing signals from the OAM-MDM system can be regarded as "small sample" and "large target", respectively. AffinityNet can be used to build an accurate nonlinear model using "small sample" based on few-shot learning and can predict the stochastic characteristic nonlinearity of OAM-MDM with a high level of generalization. As a result, the AffinityNet nonlinear equalizer can effectively compensate the stochastic nonlinearity in the OAM-MDM system, despite the large difference between the training and testing signals due to the stochastic nonlinear impairment. An experiment was conducted on a 400 Gbit/s transmission with four OAM modes using a pulse amplitude modulation-8 (PAM-8) signal over a 2 km ring-core fiber (RCF). Our experimental results show that the proposed nonlinear equalizer outperformed the conventional Volterra equalizer with improvements in receiver sensitivity of 1.7, 1.8, 3, and 3.3 dB for the four OAM modes at the 15% forward error correction (FEC) threshold, respectively. In addition, the proposed equalizer outperformed a convolutional neural network (CNN) equalizer with improvements in receiver sensitivity of 0.8, 0.5, 0.9, and 1.4 dB for the four OAM modes at the 15% FEC threshold. In the experiment, a complexity reduction of 37% and 83% of the AffinityNet equalizer is taken compared to the conventional Volterra equalizer and CNN equalizer, respectively. The proposed equalizer is a promising candidate for a high-speed OAM-MDM optical fiber communication system.

Hardware-Efficient Duobinary Neural Network Equalizers for 800 Gb/s IM/DD PAM4 Transmission Over 10 km SSMF

In this article, we discuss challenges and options for scaling IM/DD transceivers towards 800Gbps. Our focus is CWDM4 PAM4 transmission and our target distance is 10km in O-band, which is a most urgent use case for next generation optical short reach systems like data centre interconnects and networks. <p xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">At this reach and rate, chromatic dispersion (CD) becomes the main challenge. Its mitigation is essential and primarily done with digital signal processing. State of the art techniques, however, make transceivers quickly too complex. <p xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">We show upon measurement results how neural network equalization can meet Volterra equalization performance with 30% less hardware multiplier complexity. When also applying magnitude weight pruning, an additional 43% reduction is possible without performance loss across all CWDM4 lanes. If needed, an added MLSE stage can further push performance in both cases. <p xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">In any of these configurations, a key enabler against strong CD penalties is duobinary training, which is applicable to all feed-forward equalization architectures.

4Gbaud PS-16QAM D-Band Fiber-Wireless Transmission over 4.6 km by Using Balance Complex-Valued NN Equalizer with Random Oversampling.

D-band (110-170 GHz) is a promising direction for the future of 6th generation mobile networks (6G) for high-speed mobile communication since it has a large available bandwidth, and it can provide a peak rate of hundreds of Gbit/s. Compared with the traditional electrical approach, photonics millimeter wave (mm-wave) generation in D-band is more practical and effectively overcomes the bottleneck of electrical devices. However, long-distance D-band wireless transmission is still limited by some key factors such as large absorption loss and nonlinear noises. Deep neural network algorithms are regarded as an important technique to model the nonlinear wireless behavior, among which the study on complex-value equalization is critical, especially in coherent detection systems. Moreover, probabilistic shaping is useful to improve the transmission capacity but also causes an imbalanced machine learning issue. In this paper, we propose a novel complex-valued neural network equalizer coupled with balanced random oversampling (ROS). Thanks to the adaptive deep learning method for probabilistic shaping-quadrature amplitude modulation (PS-QAM), we successfully realize a 135 GHz 4Gbaud PS-16QAM with a shaping entropy of 3.56 bit/symbol wireless transmission over 4.6 km. The bit error ratio (BER) of 4Gbaud PS-16QAM can be decreased to a soft-decision forward error correction (SD-FEC) with a 25% overhead of 2 × 10-2. Therefore, we can achieve a net rate of an 11.4 Gbit/s D-band radio-over-fiber (ROF) delivery over 4.6 km air free wireless distance.

Reduced Complexity Neural Network Equalizers for Two-Dimensional Magnetic Recording

This article investigates the reduced complexity neural network (NN)-based architectures for equalization over the two-dimensional magnetic recording (TDMR) digital communication channel for data storage. We use realistic waveforms measured from a hard disk drive (HDD) with TDMR technology. We show that the multilayer perceptron (MLP) nonlinear equalizer achieves a 10.91% reduction in bit error rate (BER) over the linear equalizer with cross-entropy (CE)-based optimization. However, the MLP equalizer’s complexity is <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$6.6\times $ </tex-math></inline-formula> the linear equalizer’s complexity. Thus, we propose the reduced complexity MLP (RC-MLP) equalizers. Each RC-MLP variant consists of finite-impulse response (FIR) filters, a nonlinear activation, and a hidden delay line. A proposed RC-MLP variant entails only <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1.59\times $ </tex-math></inline-formula> the linear equalizer’s complexity while achieving a 8.23% reduction in BER over the linear equalizer.

Real-Valued Neural Network Nonlinear Equalization for Long-Reach PONs Based on SSB Modulation

Real-valued neural network nonlinear equalization for complex-valued signals combined with single sideband (SSB) modulation is proposed to be a promising scheme for long-reach passive optical networks (LR-PONs). With the assist of neural network equalizer (NNE), fiber nonlinearity induced by large launch power can be mitigated. We experimentally demonstrate a transmission of 120 Gbps 16-QAM SSB signal in C-band over an 80-km single mode fiber (SMF) without optical inline amplifier and pre-amplifier.

Electronic dispersion compensation and functional link neural network equalization for IM/DD links

Short-reach fiber optical links employing intensity modulation (IM) at the transmitter (Tx) and direct detection (DD) at the receiver (Rx), suffer from linear and nonlinear sources of impairments due to the interaction of chromatic dispersion (CD) with DD. In this article, joint electronic dispersion compensation (EDC) at the Tx, using two distinct Gerchberg–Saxton(GS) based approaches, and at the Rx, using a functional link neural network (FLNN) equalizer is demonstrated for IM/DD transmission. The first Tx approach utilizes the modified iterative GS algorithm, which partially mitigate linear and nonlinear sources of inter-symbol interference (ISI). The second Tx pre-EDC approach only pre-compensates for the linear power fading effect through implementing a GS based finite impulse response (FIR) filter. At the Rx, a T/2-spaced or T-spaced adaptive post-feed forward equalizer (FFE) is employed to fully compensate residual chromatic dispersion prior to attempting nonlinear equalization. Furthermore, a Volterra nonlinear equalizer (VNLE) is introduced to benchmark the performance and complexity of the FLNN, Subsequently, either a FLNN or a VNLE are utilized for nonlinear system identification and subsequent post-equalization mitigating uncompensated nonlinear sources of ISI. The FLNN nonlinear taps resulting from the functional expansion block are optimized using the recursive least square (RLS) algorithm. The Tx-FIR and Rx-FLNN enable 112 Gbit/s non-return to zero (NRZ) on–off keying (OOK) transmission over 20 km of single mode fiber (SMF) and 112 Gbit/s 4-level pulse-amplitude modulation (PAM-4) transmission over 10 km of SMF. It is shown that the use of Tx pre-EDC reduces the complexity of the required equalization at the Rx. In addition, the FLNN was found to offer a 74% reduction in computational complexity relative to the third-order VNLE.

Adaptive Bayesian neural networks nonlinear equalizer in a 300-Gbit/s PAM8 transmission for IM/DD OAM mode division multiplexing.

The strong stochastic nonlinear impairment induced by random mode coupling appears to be a long-standing performance-limiting problem in the orbital angular momentum (OAM) mode division multiplexing (MDM) of intensity modulation direct detection (IM/DD) transmission systems. In this Letter, we propose a Bayesian neural network (BNN) nonlinear equalizer for an OAM-MDM IM/DD transmission with three modes. Unlike conventional Volterra and convolutional neural network (CNN) equalizers with fixed weight coefficients, the weights and biases of the BNN nonlinear equalizer are regarded as probability distributions, which can accurately match the stochastic nonlinear model of the OAM-MDM. The BNN nonlinear equalizer is capable of adaptively updating its weights and biases sample-by-sample, according to the probability distribution. An experiment was conducted on a 300-Gbit/s PAM8 signal with three modes over a 2.6-km OAM-MDM RCF transmission. The experimental results demonstrate that the proposed BNN nonlinear equalizer exhibits promising solutions to effectively mitigate nonlinear distortions, which outperforms conventional Volterra and CNN equalizers with receiver sensitivity improvements of 1.0 dBm and 2.5 dBm, respectively, under hard-decision forward error correction (HD-FEC) thresholds. Moreover, compared with the Volterra and CNN equalizers, the complexity of the OAM-MDM is significantly improved through the BNN nonlinear equalizer. The proposed BNN nonlinear equalizer is a promising candidate for the high capacity inter-data center interconnects.

Probabilistic neural network equalizer for nonlinear mitigation in OAM mode division multiplexed optical fiber communication.

Orbital angular momentum (OAM) mode-division multiplexing (MDM) is a key technique to achieve ultra-high-capacity optical fiber communications. However, the high nonlinear impairment from optoelectronic devices, such as spatial light modulators, modulators, and photodiodes, is a long-standing challenge for OAM-MDM. In this paper, an equalizer based on a probabilistic neural network (PNN) is presented to mitigate the nonlinear impairment for an OAM-MDM fiber communication system with 32 GBaud Nyquist pulse amplitude modulation-8 (PAM8) intensity-modulation direct-detection (IM-DD) signals. PNN equalizer can calculate the distribution of the nonlinearity using Bayesian decision theory and thus mitigate the stochastic nonlinear impairment of the received signal. Experimental results show that compared with the convolutional neural network (CNN) equalizer, the PNN equalizer improves the receiver sensitivity by 0.6dB and 2dB for two OAM modes with l = + 3 and l = + 4 at the 20% FEC limit, respectively. Moreover, compared with Volterra or CNN equalizers, the PNN equalizer can reduce the computation complexity significantly, which has great potential to mitigate the nonlinear signal distortions in high-speed IM-DD OAM-MDM fiber communication systems.

Volterra-Aided Neural Network Equalization for Channel Impairment Compensation in Visible Light Communication System

This paper addresses the channel impairment to enhance the system performance of visible light communication (VLC). Inspired by the model-solving procedure in the conventional equalizer, the channel impairment compensation is formulated as a spatial memory pattern prediction problem, then we propose efficient deep-learning (DL)-based nonlinear post-equalization, combining the Volterra-aided convolutional neural network (CNN) and long-short term memory (LSTM) neural network, to mitigate the system nonlinearity and then recover the original transmitted signal from the distorted one at the receiver end. The Volterra structure is employed to construct a spatial pattern that can be easily interpreted by the proposed scheme. Then, we take advantage of the CNN to extract the implicit feature of channel impairments and utilize the LSTM to predict the memory sequence. Results demonstrate that the proposed scheme can provide a fairly fast convergence during the training stage and can effectively mitigate the overall nonlinearity of the system at testing. Furthermore, it can recover the original signal accurately and exhibits an excellent bit error rate performance as compared with the conventional equalizer, demonstrating the prospect and validity of this methodology for channel impairment compensation.

Low-complexity Volterra-inspired neural network equalizer in 100-G band-limited IMDD PON system.

One of the most promising solutions for 100 Gb/s line-rate passive optical networks (PONs) is intensity modulation and direct detection (IMDD) technology together with a digital signal processing- (DSP-) based equalizer for its advantages of system simplicity, cost-effectiveness, and energy-efficiency. However, due to restricted hardware resources, the effective neural network (NN) equalizer and Volterra nonlinear equalizer (VNLE) have the drawback of high implementation complexity. In this paper, we incorporate an NN with the physical principles of a VNLE to construct a white-box low-complexity Volterra-inspired neural network (VINN) equalizer. This equalizer has better performance than a VNLE at the same complexity and attains similar performance with much lower complexity than a VNLE with optimized structural hyperparameter. The effectiveness of the proposed equalizer is verified in 1310 nm band-limited IMDD PON systems. A 30.5-dB power budget is achieved with the 10-G-class transmitter.

Optimizations and investigations for transfer learning of iteratively pruned neural network equalizers for data center networking.

In this work, for the first time to the best of our knowledge, we introduce the iterative pruning technique into the transfer learning (TL) of neural network equalizers (NNE) deployed in optical links with different length. For the purpose of time saving during the training period of NNE, TL migrates the NNE parameters which have been already trained on the source link to the newly-routed link (the target link), which has been proved to outperform the training initialized with the random state. Based on simulations, we proved that iterative pruning technique could further enhance the convergence speed during TL between the source and target links. Moreover, we quantitatively investigate the marginal effects of pruned threshold and pruned span on the convergence performance in various transmission distance scenarios. In addition, we observed a trade-off between performance stability and complexity of NNE, which requires to be optimized compromisingly by choosing an appropriate equalizer scale.

Low Complexity Neural Network Equalization Based on Multi-Symbol Output Technique for 200+ Gbps IM/DD Short Reach Optical System

Nowadays, Neural network (NN) has been proved to be an effective solution for nonlinear equalization in short reach optical systems. However, recent research has mainly focused on implementing more powerful NNs for equalization, while ignoring their adaptability to equalization tasks. In this paper, we propose Multi-Symbol Output (MSO)-Neural Networks (NN) for nonlinear equalization in high-speed short reach optical interconnects. The results show that the proposed MSO design works well on Deep Neural Networks (DNN), Long Short-Term Memory neural networks (LSTM) and Gate Recurrent Unit (GRU), which are the recent NN-based equalization structures. By increasing output symbols of the NNs, the number of slide windows in equalization can be sharply reduced, and so the complexity is reduced. By the same time, more information is brought to the MSO-NNs in back-propagations, therefore performance gain achieved. A 212-Gb/s 1-km Pulse Amplitude Modulation (PAM)-4 optical link is experimentally demonstrated as the target system, and the proposed MSO-NN equalizers are used to compare with traditional equalization algorithms including Volterra Nonlinear Equalizer (VNE) and single-symbol output NNs. Experimental results show that MSO design could help reduce the complexity of NN required for nonlinear equalization in the target system by around 2/3, and the proposed MSO-LSTM performs much better than VNE and 1 dB better than SSO-LSTM at the same time. Based on the proposed MSO-LSTM, transmission with BER under HD-FEC over 1 km NZDSF is achieved with a ROP at -2 dBm. Our work is well expandable and the proposed MSO design can be extended to other NN-based equalizers, which can help reduce complexity and learn more info from the training data and gain performance. The proposed MSO-NNs further enhance the performance and reduces the complexity of NN equalizers, provides assurance for future real-time high-speed short-range optical systems, and brings new ideas to NN-based equalizer design.

Neural Network Equalizer in Visible Light Communication: State of the Art and Future Trends

As 6G research progresses, both visible light communication (VLC) and artificial intelligence (AI) become important components, which makes them appear to converge. Neural networks (NN) as equalizers are gradually occupying an increasingly important position in the research of the physical layer of VLC, especially in nonlinear compensation. In this paper, we will propose three categories of neural network equalizers, including input data reconfiguration NN, network reconfiguration NN and loss function reconfiguration NN. We give the definitions of these three neural networks and their applications in VLC systems. This work allows the reader to have a clearer understanding and future trends of neural networks in visible light communication, especially in terms of equalizers.

Recurrent Neural Network Equalization for Wireline Communication Systems

Satisfying the demand for ever increasing data rates in wireline communication systems is challenging due to high frequency-dependent channel insertion loss causing significant inter-symbol interference (ISI). This motivates the use of ADC-based receiver architectures that allow for digital equalizer implementations. However, the large tap count feed-forward equalizers (FFEs) required at high data rates causes noise enhancement that can degrade signal-to-noise ratio (SNR) and high complexity is required to implement multi-tap decision feedback equalizers (DFEs). Moreover, analog front-end (AFE) nonlinearity can limit the performance of these traditional FFE-DFE architectures. This brief investigates an alternative approach with recurrent neural network (RNN) equalization based on long short-term memory (LSTM) and gate recurrent unit (GRU) cells. 64 Gb/s PAM4 modeling results show that the RNN topologies offer significant BER improvement over a wide SNR range for a channel with 30 dB loss at 16 GHz and enable operation over a channel with a large notch at 10 GHz that the traditional FFE-DFE topology cannot support. Synthesis results utilizing a 22 nm FinFET process yield only a 34% and 29% increase in receiver power consumption for the LSTM and GRU implementations, respectively, to enable this improved performance.

Improved Recording Performance With Writer Design and Joint Learning-Based Neural Network for Heat-Assisted Interlaced Magnetic Recording

Heat-assisted interlaced magnetic recording (HIMR) further increases the areal density with the interlaced track layout architecture compared to conventional heat-assisted magnetic recording. However, besides the 2-D inter-symbol interference (ISI) and thermal jitters, the transition curvatures due to the circular thermal profile cause severe erasures after write and obvious nonlinear distortions of readback signals. This deteriorates the recording performance severely. Especially the degrees of transition curvatures for the low temperature written track (LT) and high temperature written track (HT) are distinct, which makes the correction of transition curvature with only the writer design impossible. Correspondingly, a comprehensive method consisting of writer design and multitrack detection algorithm with joint learning-based neural network (JLNN) is proposed to mitigate the effects of transition curvature and 2-D ISI for both LT and HT of HIMR. First, a designed writer with splitting main poles produces the spatially nonuniform write field along the crosstrack direction to correct the transition curvature of LT. Then the multitrack detection with JLNN and improved Bahl–Cocke–Jelinek–Raviv (BCJR) detector (JLNN-IB) algorithm is studied to further mitigate the effects of residual transition curvatures, thermal jitters, and 2-D ISI for both HT and LT. Specifically, the readback signals of read head array and decoded log-likelihood ratio (LLR) outputted by the low-density parity check (LDPC) decoder are fed into JLNN to obtain the equalized signal and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$a$ </tex-math></inline-formula> <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">priori</i> probability (AP) of three tracks. Then the equalized signal as well as AP of current track and corresponding soft bit estimations of adjacent tracks are embedded into the branch metrics of improved BCJR detector for data detection, and the LDPC decoder is cascaded for the error correction. The simulation indicates that at the recording density of 2.11 Tb/in <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> , the proposed comprehensive writer design and JLNN-IB algorithm provide 9.2 and 9.7 dB signal to noise ratio (SNR) gains compared to the conventional 2-D neural network equalizer and 2-D linear equalizer under the uniform write field and circular thermal profile, respectively.

Low-Complexity Recurrent Neural Network Based Equalizer With Embedded Parallelization for 100-Gbit/s/λ PON

To meet the demand of emerging applications, such as fixed-mobile convergence for the fifth generation of mobile networks and beyond, a 100-Gbit/s/λ access network becomes the next priority for the passive optical network roadmap. We experimentally demonstrate the transmission of 100-Gbit/s/λ intensity modulation and direct detection passive optical network based on four-level pulsed amplitude modulation in the O-band by using 25G-class optics. To mitigate the severe distortions caused by inter-symbol interference and fiber nonlinearity, a low-complexity recurrent neural network based equalizer with parallel outputs is proposed. Experimental results show that the proposed recurrent neural network equalizer can consistently outperform fully-connected neural network with the same input/output size and number of training parameters. The neural network equalizer's sensitivity against quantization is also evaluated. To further understand the complexity and actual hardware resource consumption of the parallel-output equalizers, we implement an 8bits-integer-quantized neural network model using FPGA, with the benefits and challenges validated and discussed.