Search

Intracranial hemorrhages in head CT scans serve as a first line tool to help specialists diagnose different types. However, their types have diverse shapes in the same type but similar confusing shape, size and location between types. To solve this problem, this paper proposes an all attention U-Net. It uses channel attentions in the U-Net encoder side to enhance class specific feature extraction, and space and channel attentions in the U-Net decoder side for more accurate shape extraction and type classification. The simulation results show up to a 31.8\% improvement compared to baseline, ResNet50 + U-Net, and better performance than in cases with limited attention., Comment: 2022 IEEE Biomedical Circuits and Systems Conference (BioCAS)

Image prefiltering with just noticeable distortion (JND) improves coding efficiency in a visual lossless way by filtering the perceptually redundant information prior to compression. However, real JND cannot be well modeled with inaccurate masking equations in traditional approaches or image-level subject tests in deep learning approaches. Thus, this paper proposes a fine-grained JND prefiltering dataset guided by image quality assessment for accurate block-level JND modeling. The dataset is constructed from decoded images to include coding effects and is also perceptually enhanced with block overlap and edge preservation. Furthermore, based on this dataset, we propose a lightweight JND prefiltering network, IQNet, which can be applied directly to different quantization cases with the same model and only needs 3K parameters. The experimental results show that the proposed approach to Versatile Video Coding could yield maximum/average bitrate savings of 41\%/15\% and 53\%/19\% for all-intra and low-delay P configurations, respectively, with negligible subjective quality loss. Our method demonstrates higher perceptual quality and a model size that is an order of magnitude smaller than previous deep learning methods.

Low power deep learning accelerators on the speech processing enable real-time applications on edge devices. However, most of the existing accelerators suffer from high power consumption and focus on image applications only. This paper presents a low power accelerator for speech separation through algorithm and hardware optimizations. At the algorithm level, the model is compressed with structured sensitivity as well as unstructured pruning, and further quantized to the shifted 8-bit floating-point format instead of the 32-bit floating-point format. The computations with the zero kernel and zero activation values are skipped by decomposition of the dilated and transposed convolutions. At the hardware level, the compressed model is then supported by an architecture with eight independent multipliers and accumulators (MACs) with a simple zero-skipping hardware to take advantage of the activation sparsity and low power processing. The proposed approach reduces the model size by 95.44\% and computation complexity by 93.88\%. The final implementation with the TSMC 40 $nm$ process can achieve real-time speech separation and consumes 1.6 mW power when operated at 150 MHz. The normalized energy efficiency and area efficiency are 2.344 TOPS/W and 14.42 GOPS/mm$^2$, respectively.

Deep-learning accelerators are increasingly in demand; however, their performance is constrained by the size of the feature map, leading to high bandwidth requirements and large buffer sizes. We propose an adaptive scale feature map compression technique leveraging the unique properties of the feature map. This technique adopts independent channel indexing given the weak channel correlation and utilizes a cubical-like block shape to benefit from strong local correlations. The method further optimizes compression using a switchable endpoint mode and adaptive scale interpolation to handle unimodal data distributions, both with and without outliers. This results in 4$\times$ and up to 7.69$\times$ compression rates for 16-bit data in constant and variable bitrates, respectively. Our hardware design minimizes area cost by adjusting interpolation scales, which facilitates hardware sharing among interpolation points. Additionally, we introduce a threshold concept for straightforward interpolation, preventing the need for intricate hardware. The TSMC 28nm implementation showcases an equivalent gate count of 6135 for the 8-bit version. Furthermore, the hardware architecture scales effectively, with only a sublinear increase in area cost. Achieving a 32$\times$ throughput increase meets the theoretical bandwidth of DDR5-6400 at just 7.65$\times$ the hardware cost.

Deep learning-driven superresolution (SR) outperforms traditional techniques but also faces the challenge of high complexity and memory bandwidth. This challenge leads many accelerators to opt for simpler and shallow models like FSRCNN, compromising performance for real-time needs, especially for resource-limited edge devices. This paper proposes an energy-efficient SR accelerator, ACNPU, to tackle this challenge. The ACNPU enhances image quality by 0.34dB with a 27-layer model, but needs 36\% less complexity than FSRCNN, while maintaining a similar model size, with the \textit{decoupled asymmetric convolution and split-bypass structure}. The hardware-friendly 17K-parameter model enables \textit{holistic model fusion} instead of localized layer fusion to remove external DRAM access of intermediate feature maps. The on-chip memory bandwidth is further reduced with the \textit{input stationary flow} and \textit{parallel-layer execution} to reduce power consumption. Hardware is regular and easy to control to support different layers by \textit{processing elements (PEs) clusters with reconfigurable input and uniform data flow}. The implementation in the 40 nm CMOS process consumes 2333 K gate counts and 198KB SRAMs. The ACNPU achieves 31.7 FPS and 124.4 FPS for x2 and x4 scales Full-HD generation, respectively, which attains 4.75 TOPS/W energy efficiency., Comment: 9 pages, 14 figures

Modern infrared (IR) microscopy, communication, and sensing systems demand control of the spectral characteristics and polarization states of light. Typically, these systems require the cascading of multiple filters, polarization optics and rotating components to manipulate light, inevitably increasing their sizes and complexities. Here, we report two-terminal mid-infrared (mid-IR) emitters with electrically controllable spectral and polarization properties. Our devices are composed of two back-to-back p-n junctions formed by stacking anisotropic light-emitting materials, black phosphorus and black arsenic-phosphorus with MoS2. By controlling the crystallographic orientations and engineering the band profile of heterostructures, the emissions of two junctions exhibit distinct spectral ranges and polarization directions; more importantly, these two electroluminescence (EL) units can be independently activated, depending on the polarity of the applied bias. Furthermore, we show that when operating our emitter under the polarity-switched pulse mode, its EL exhibits the characteristics of broad spectral coverage, encompassing the entire first mid-IR atmospheric window, and electrically tunable spectral shapes. Our results provide the basis for developing groundbreaking technology in the field of light emitters.

Stacking two-dimensional layered materials such as graphene and transitional metal dichalcogenides with nonzero interlayer twist angles has recently become attractive because of the emergence of novel physical properties. Stacking of one-dimensional nanomaterials offers the lateral stacking offset as an additional parameter for modulating the resulting material properties. Here, we report that the edge states of twisted bilayer zigzag graphene nanoribbons (TBZGNRs) can be tuned with both the twist angle and the stacking offset. Strong edge state variations in the stacking region are first revealed by density functional theory (DFT) calculations. We construct and characterize twisted bilayer zigzag graphene nanoribbon (TBZGNR) systems on a Au(111) surface using scanning tunneling microscopy. A detailed analysis of three prototypical orthogonal TBZGNR junctions exhibiting different stacking offsets by means of scanning tunneling spectroscopy reveals emergent near-zero-energy states. From a comparison with DFT calculations, we conclude that the emergent edge states originate from the formation of flat bands whose energy and spin degeneracy are highly tunable with the stacking offset. Our work highlights fundamental differences between 2D and 1D twistronics and spurs further investigation of twisted one-dimensional systems., Comment: 42 pages, 18 figures

Previous gait phase detection as convolutional neural network (CNN) based classification task requires cumbersome manual setting of time delay or heavy overlapped sliding windows to accurately classify each phase under different test cases, which is not suitable for streaming Inertial-Measurement-Unit (IMU) sensor data and fails to adapt to different scenarios. This paper presents a segmentation based gait phase detection with only a single six-axis IMU sensor, which can easily adapt to both walking and running at various speeds. The proposed segmentation uses CNN with gait phase aware receptive field setting and IMU oriented processing order, which can fit to high sampling rate of IMU up to 1000Hz for high accuracy and low sampling rate down to 20Hz for real time calculation. The proposed model on the 20Hz sampling rate data can achieve average error of 8.86 ms in swing time, 9.12 ms in stance time and 96.44\% accuracy of gait phase detection and 99.97\% accuracy of stride detection. Its real-time implementation on mobile phone only takes 36 ms for 1 second length of sensor data., Comment: 5 pages, 5 figures, published in IEEE ISCAS 2020

This paper presents a convolutional neural network based foot motion tracking with only six-axis Inertial-Measurement-Unit (IMU) sensor data. The presented approach can adapt to various walking conditions by adopting differential and window based input. The training data are further augmented by sliding and random window samplings on IMU sensor data to increase data diversity for better performance. The proposed approach fuses predictions of three dimensional output into one model. The proposed fused model can achieve average error of 2.30+-2.23 cm in X-axis, 0.91+-0.95 cm in Y-axis and 0.58+-0.52 cm in Z-axis., Comment: 4 pages, 4 figures

Keyword spotting has gained popularity as a natural way to interact with consumer devices in recent years. However, because of its always-on nature and the variety of speech, it necessitates a low-power design as well as user customization. This paper describes a low-power, energy-efficient keyword spotting accelerator with SRAM based in-memory computing (IMC) and on-chip learning for user customization. However, IMC is constrained by macro size, limited precision, and non-ideal effects. To address the issues mentioned above, this paper proposes bias compensation and fine-tuning using an IMC-aware model design. Furthermore, because learning with low-precision edge devices results in zero error and gradient values due to quantization, this paper proposes error scaling and small gradient accumulation to achieve the same accuracy as ideal model training. The simulation results show that with user customization, we can recover the accuracy loss from 51.08\% to 89.76\% with compensation and fine-tuning and further improve to 96.71\% with customization. The chip implementation can successfully run the model with only 14$uJ$ per decision. When compared to the state-of-the-art works, the presented design has higher energy efficiency with additional on-chip model customization capabilities for higher accuracy., Comment: 10 pages, 18 figures, to be published in IEEE Transaction on VLSI, 2022

Following the success of the natural language processing, the transformer for vision applications has attracted significant attention in recent years due to its excellent performance. However, existing deep learning hardware accelerators for vision cannot execute this structure efficiently due to significant model architecture differences. As a result, this paper proposes the hardware accelerator for vision transformers with row-wise scheduling, which decomposes major operations in vision transformers as a single dot product primitive for a unified and efficient execution. Furthermore, by sharing weights in columns, we can reuse the data and reduce the usage of memory. The implementation with TSMC 40nm CMOS technology only requires 262K gate count and 149KB SRAM buffer for 403.2 GOPS throughput at 600MHz clock frequency., Comment: 5 pages, 6 figures, published in IEEE AICAS 2022

Deep learning based superresolution achieves high-quality results, but its heavy computational workload, large buffer, and high external memory bandwidth inhibit its usage in mobile devices. To solve the above issues, this paper proposes a real-time hardware accelerator with the tilted layer fusion method that reduces the external DRAM bandwidth by 92\% and just needs 102KB on-chip memory. The design implemented with a 40nm CMOS process achieves 1920x1080@60fps throughput with 544.3K gate count when running at 600MHz; it has higher throughput and lower area cost than previous designs., Comment: 5 pages, 6 figures, published in ISCAS 2022

In-memory computing is becoming a popular architecture for deep-learning hardware accelerators recently due to its highly parallel computing, low power, and low area cost. However, in-RRAM computing (IRC) suffered from large device variation and numerous nonideal effects in hardware. Although previous approaches including these effects in model training successfully improved variation tolerance, they only considered part of the nonideal effects and relatively simple classification tasks. This paper proposes a joint hardware and software optimization strategy to design a hardware-robust IRC macro for object detection. We lower the cell current by using a low word-line voltage to enable a complete convolution calculation in one operation that minimizes the impact of nonlinear addition. We also implement ternary weight mapping and remove batch normalization for better tolerance against device variation, sense amplifier variation, and IR drop problem. An extra bias is included to overcome the limitation of the current sensing range. The proposed approach has been successfully applied to a complex object detection task with only 3.85\% mAP drop, whereas a naive design suffers catastrophic failure under these nonideal effects., Comment: 10 pages, 18 figures

Stride length estimation using inertial measurement unit (IMU) sensors is getting popular recently as one representative gait parameter for health care and sports training. The traditional estimation method requires some explicit calibrations and design assumptions. Current deep learning methods suffer from few labeled data problem. To solve above problems, this paper proposes a single convolutional neural network (CNN) model to predict stride length of running and walking and classify the running or walking type per stride. The model trains its pretext task with self-supervised learning on a large unlabeled dataset for feature learning, and its downstream task on the stride length estimation and classification tasks with supervised learning with a small labeled dataset. The proposed model can achieve better average percent error, 4.78\%, on running and walking stride length regression and 99.83\% accuracy on running and walking classification, when compared to the previous approach, 7.44\% on the stride length estimation., Comment: 8 pages, 11 figures

Hardware accelerator for convolution neural network (CNNs) enables real time applications of artificial intelligence technology. However, most of the accelerators only support dense CNN computations or suffers complex control to support fine grained sparse networks. To solve above problem, this paper presents an efficient CNN accelerator with 1-D vector broadcasted input to support both dense network as well as vector sparse network with the same hardware and low overhead. The presented design achieves 1.93X speedup over the dense CNN computations., Comment: 5 pages, 13 figures, published in IEEE ISCAS 2019

Hardware accelerators for convolution neural networks (CNNs) enable real-time applications of artificial intelligence technology. However, most of the existing designs suffer from low hardware utilization or high area cost due to complex dataflow. This paper proposes a hardware efficient vectorwise CNN accelerator that adopts a 3$\times$3 filter optimized systolic array using 1-D broadcast dataflow to generate partial sum. This enables easy reconfiguration for different kinds of kernels with interleaved input or elementwise input dataflow. This simple and regular data flow results in low area cost while attains high hardware utilization. The presented design achieves 99\%, 97\%, 93.7\%, 94\% hardware utilization for VGG-16, ResNet-34, GoogLeNet, and Mobilenet, respectively. Hardware implementation with TSMC 40nm technology takes 266.9K NAND gate count and 191KB SRAM to support 168GOPS throughput and consumes only 154.98mW when running at 500MHz operating frequency, which has superior area and power efficiency than other designs., Comment: 11 pages, 21 figures

Memory bandwidth has become the real-time bottleneck of current deep learning accelerators (DLA), particularly for high definition (HD) object detection. Under resource constraints, this paper proposes a low memory traffic DLA chip with joint hardware and software optimization. To maximize hardware utilization under memory bandwidth, we morph and fuse the object detection model into a group fusion-ready model to reduce intermediate data access. This reduces the YOLOv2's feature memory traffic from 2.9 GB/s to 0.15 GB/s. To support group fusion, our previous DLA based hardware employes a unified buffer with write-masking for simple layer-by-layer processing in a fusion group. When compared to our previous DLA with the same PE numbers, the chip implemented in a TSMC 40nm process supports 1280x720@30FPS object detection and consumes 7.9X less external DRAM access energy, from 2607 mJ to 327.6 mJ., Comment: 11 pages, 14 figures, to be published IEEE Transactions on Very Large Scale Integration (VLSI) Systems

Semantic outdoor scene understanding based on 3D LiDAR point clouds is a challenging task for autonomous driving due to the sparse and irregular data structure. This paper takes advantages of the uneven range distribution of different LiDAR laser beams to propose a range aware instance segmentation network, RangeSeg. RangeSeg uses a shared encoder backbone with two range dependent decoders. A heavy decoder only computes top of a range image where the far and small objects locate to improve small object detection accuracy, and a light decoder computes whole range image for low computational cost. The results are further clustered by the DBSCAN method with a resolution weighted distance function to get instance-level segmentation results. Experiments on the KITTI dataset show that RangeSeg outperforms the state-of-the-art semantic segmentation methods with enormous speedup and improves the instance-level segmentation performance on small and far objects. The whole RangeSeg pipeline meets the real time requirement on NVIDIA\textsuperscript{\textregistered} JETSON AGX Xavier with 19 frames per second in average., Comment: 10 pages, 7 figures. IEEE Transactions on Intelligent Vehicles (2021)

Computing-in-memory (CIM) has attracted significant attentions in recent years due to its massive parallelism and low power consumption. However, current CIM designs suffer from large area overhead of small CIM macros and bad programmablity for model execution. This paper proposes a programmable CIM processor with a single large sized CIM macro instead of multiple smaller ones for power efficient computation and a flexible instruction set to support various binary 1-D convolution Neural Network (CNN) models in an easy way. Furthermore, the proposed architecture adopts the pooling write-back method to support fused or independent convolution/pooling operations to reduce 35.9\% of latency, and the flexible ping-pong feature SRAM to fit different feature map sizes during layer-by-layer execution.The design fabricated in TSMC 28nm technology achieves 150.8 GOPS throughput and 885.86 TOPS/W power efficiency at 10 MHz when executing our binary keyword spotting model, which has higher power efficiency and flexibility than previous designs., Comment: 5 pages, 7 figures, published in IEEE ISCAS 2022

Hardware acceleration for dilated and transposed convolution enables real time execution of related tasks like segmentation, but current designs are specific for these convolutional types or suffer from complex control for reconfigurable designs. This paper presents a design that decomposes input or weight for dilated and transposed convolutions respectively to skip redundant computations and thus executes efficiently on existing dense CNN hardware as well. The proposed architecture can cut down 87.8\% of the cycle counts to achieve 8.2X speedup over a naive execution for the ENet case., Comment: 10 pages, 12 figures, published in IEEE ISCAS 2020

Gait phase detection with convolution neural network provides accurate classification but demands high computational cost, which inhibits real time low power on-sensor processing. This paper presents a segmentation based gait phase detection with a width and depth downscaled U-Net like model that only needs 0.5KB model size and 67K operations per second with 95.9% accuracy to be easily fitted into resource limited on sensor microcontroller., Comment: 2 pages, 2 figures, published in 2020 ICCE-TW

With the popularity of the deep neural network (DNN), hardware accelerators are demanded for real time execution. However, lengthy design process and fast evolving DNN models make hardware evaluation hard to meet the time to market need. This paper proposes a pre-RTL DNN hardware evaluator that supports conventional layer-by-layer processing as well as the fused layer processing for low external bandwidth requirement. The evaluator supports two state-of-the-art accelerator architectures and finds the best hardware and layer fusion group The experimental results show the layer fusion scheme can achieve 55.6% memory bandwidth reduction, 36.7% latency improvement and 49.2% energy reduction compared with layer-by-layer operation., Comment: 2 pages, 2 figures, published in IEEE ISOCC 2021

There is a great interest in many approaches towards blockchain in providing a solution to record transactions in a decentralized way. However, there are some limitations when storing large files or documents on the blockchain. In order to meet the requirements of storing relatively large data, a decentralized storage medium is produced. IPFS is a distributed file system which is content-addressable. It works very similar to the blockchain network. There are some attempts which take advantage of the blockchain concept and IPFS to design new approaches. Unfortunately, there are some inefficiencies in sharing data using the combination of IPFS and blockchain. In this paper, we proposed a secure file sharing system that brings a distributed access control and group key management by the adoption of the IPFS proxy. The IPFS proxy which plays an important role in the design is adopted to take responsibility for the control policies. The combination of the IPFS server and the blockchain network with the adoption of the IPFS proxy make a secure file sharing system which the members on the system can create new groups or join different groups by their own choice. Although there is no access control mechanism in IPFS server and blockchain network, the secure file sharing system manages the access control policies. The members access files only belong to the group they authorized., Comment: 4 pages, 4 figures, published in Proceedings of the 2020 2nd International Electronics Communication Conference

The large amount of memory bandwidth between local buffer and external DRAM has become the speedup bottleneck of CNN hardware accelerators, especially for activation maps. To reduce memory bandwidth, we propose to learn pruning unimportant blocks dynamically with zero block regularization of activation maps (Zebra). This strategy has low computational overhead and could easily integrate with other pruning methods for better performance. The experimental results show that the proposed method can reduce 70\% of memory bandwidth for Resnet-18 on Tiny-Imagenet within 1\% accuracy drops and 2\% accuracy gain with the combination of Network Slimming., Comment: 5 pages, 5 figures, published in IEEE ISCAS 2021

Spiking neural networks (SNNs) that enable low-power design on edge devices have recently attracted significant research. However, the temporal characteristic of SNNs causes high latency, high bandwidth and high energy consumption for the hardware. In this work, we propose a binary weight spiking model with IF-based Batch Normalization for small time steps and low hardware cost when direct training with input encoding layer and spatio-temporal back propagation (STBP). In addition, we propose a vectorwise hardware accelerator that is reconfigurable for different models, inference time steps and even supports the encoding layer to receive multi-bit input. The required memory bandwidth is further reduced by two-layer fusion mechanism. The implementation result shows competitive accuracy on the MNIST and CIFAR-10 datasets with only 8 time steps, and achieves power efficiency of 25.9 TOPS/W., Comment: 5 pages, 8 figures, published in IEEE ISCAS 2021

Spiking neural networks (SNNs), which are inspired by the human brain, have recently gained popularity due to their relatively simple and low-power hardware for transmitting binary spikes and highly sparse activation maps. However, because SNNs contain extra time dimension information, the SNN accelerator will require more buffers and take longer to infer, especially for the more difficult high-resolution object detection task. As a result, this paper proposes a sparse compressed spiking neural network accelerator that takes advantage of the high sparsity of activation maps and weights by utilizing the proposed gated one-to-all product for low power and highly parallel model execution. The experimental result of the neural network shows 71.5$\%$ mAP with mixed (1,3) time steps on the IVS 3cls dataset. The accelerator with the TSMC 28nm CMOS process can achieve 1024$\times$576@29 frames per second processing when running at 500MHz with 35.88TOPS/W energy efficiency and 1.05mJ energy consumption per frame., Comment: 11 pages, 18 figures, to be published in IEEE Transactions on Circuits and Systems--I: Regular Papers

Increasingly, convolution neural network (CNN) based super resolution models have been proposed for better reconstruction results, but their large model size and complicated structure inhibit their real-time hardware implementation. Current hardware designs are limited to a plain network and suffer from lower quality and high memory bandwidth requirements. This paper proposes a super resolution hardware accelerator with hardware efficient pixel attention that just needs 25.9K parameters and simple structure but achieves 0.38dB better reconstruction images than the widely used FSRCNN. The accelerator adopts full model block wise convolution for full model layer fusion to reduce external memory access to model input and output only. In addition, CNN and pixel attention are well supported by PE arrays with distributed weights. The final implementation can support full HD image reconstruction at 30 frames per second with TSMC 40nm CMOS process., Comment: 5 pages, 5 figures, published in IEEE ISCAS 2022

The layered crystal of EuSn$_2$As$_2$ has a Bi$_2$Te$_3$-type structure in rhombohedral ($R\bar{3}m$) symmetry and has been confirmed to be an intrinsic magnetic topological insulator at ambient conditions. Combining {\it ab initio} calculations and \emph{in-situ} x-ray diffraction measurements, we identify a new monoclinic EuSn$_2$As$_2$ structure in $C2/m$ symmetry above $\sim$14 GPa. It has a three-dimensional network made up of honeycomb-like Sn sheets and zigzag As chains, transformed from the layered EuSn$_2$As$_2$ via a two-stage reconstruction mechanism with the connecting of Sn-Sn and As-As atoms successively between the buckled SnAs layers. Its dynamic structural stability has been verified by phonon mode analysis. Electrical resistance measurements reveal an insulator-metal-superconductor transition at low temperature around 5 and 15 GPa, respectively, according to the structural conversion, and the superconductivity with a \textit{T}${\rm {_C}}$ value of $\sim 4$ K is observed up to 30.8 GPa. These results establish a high-pressure EuSn$_2$As$_2$ phase with intriguing structural and electronic properties and expand our understandings about the layered magnetic topological insulators.

Light-emitting diodes (LEDs) based on III-V/II-VI materials have delivered a compelling performance in the mid-infrared (mid-IR) region, which enabled wide-ranging applications, including environmental monitoring, defense and medical diagnostics. Continued efforts are underway to realize on-chip sensors via heterogeneous integration of mid-IR emitters on a silicon photonic chip. But the uptake of such approach is limited by the high costs and interfacial strains, associated with the process of heterogeneous integrations. Here, the black phosphorus (BP)-based van der Waals (vdW) heterostructures are exploited as room temperature LEDs. The demonstrated devices can emit linearly polarized light, and their spectra cover the technologically important mid-IR atmospheric window (3-4 um). Additionally, the BP LEDs exhibit fast modulation speed as well as exceptional stability, and its peak extrinsic quantum efficiency (QE~0.9%) is comparable to the III-V/II-VI mid-IR LEDs. By leveraging the integrability of vdW heterostructures, we further demonstrate a silicon photonic waveguide-integrated BP LED. The reported hybrid platform holds great promise for mid-IR silicon photonics., Comment: 12 pages, 4 figures

Asymmetric nonlinear weight update is considered as one of the major obstacles for realizing hardware neural networks based on analog resistive synapses because it significantly compromises the online training capability. This paper provides new solutions to this critical issue through co-optimization with the hardware-applicable deep-learning algorithms. New insights on engineering activation functions and a threshold weight update scheme effectively suppress the undesirable training noise induced by inaccurate weight update. We successfully trained a two-layer perceptron network online and improved the classification accuracy of MNIST handwritten digit dataset to 87.8/94.8% by using 6-bit/8-bit analog synapses, respectively, with extremely high asymmetric nonlinearity., Comment: IEEE Journal on Emerging and Selected Topics in Circuits and Systems 2017