Future Hardware Implementation Research Articles

Evaluation of a Voltametric E-Tongue Combined with Data Preprocessing for Fast and Effective Machine Learning-Based Classification of Tomato Purées by Cultivar.

The potential of a voltametric E-tongue coupled with a custom data pre-processing stage to improve the performance of machine learning techniques for rapid discrimination of tomato purées between cultivars of different economic value has been investigated. To this aim, a sensor array with screen-printed carbon electrodes modified with gold nanoparticles (GNP), copper nanoparticles (CNP) and bulk gold subsequently modified with poly(3,4-ethylenedioxythiophene) (PEDOT), was developed to acquire data to be transformed by a custom pre-processing pipeline and then processed by a set of commonly used classifiers. The GNP and CNP-modified electrodes, selected based on their sensitivity to soluble monosaccharides, demonstrated good ability in discriminating samples of different cultivars. Among the different data analysis methods tested, Linear Discriminant Analysis (LDA) proved to be particularly suitable, obtaining an average F1 score of 99.26%. The pre-processing stage was beneficial in reducing the number of input features, decreasing the computational cost, i.e., the number of computing operations to be performed, of the entire method and aiding future cost-efficient hardware implementation. These findings proved that coupling the multi-sensing platform featuring properly modified sensors with the custom pre-processing method developed and LDA provided an optimal tradeoff between analytical problem solving and reliable chemical information, as well as accuracy and computational complexity. These results can be preliminary to the design of hardware solutions that could be embedded into low-cost portable devices.

Event-Based Optical Flow Estimation with Spatio-Temporal Backpropagation Trained Spiking Neural Network.

The advantages of an event camera, such as low power consumption, large dynamic range, and low data redundancy, enable it to shine in extreme environments where traditional image sensors are not competent, especially in high-speed moving target capture and extreme lighting conditions. Optical flow reflects the target's movement information, and the target's detailed movement can be obtained using the event camera's optical flow information. However, the existing neural network methods for optical flow prediction of event cameras has the problems of extensive computation and high energy consumption in hardware implementation. The spike neural network has spatiotemporal coding characteristics, so it can be compatible with the spatiotemporal data of an event camera. Moreover, the sparse coding characteristic of the spike neural network makes it run with ultra-low power consumption on neuromorphic hardware. However, because of the algorithmic and training complexity, the spike neural network has not been applied in the prediction of the optical flow for the event camera. For this case, this paper proposes an end-to-end spike neural network to predict the optical flow of the discrete spatiotemporal data stream for the event camera. The network is trained with the spatio-temporal backpropagation method in a self-supervised way, which fully combines the spatiotemporal characteristics of the event camera while improving the network performance. Compared with the existing methods on the public dataset, the experimental results show that the method proposed in this paper is equivalent to the best existing methods in terms of optical flow prediction accuracy, and it can save 99% more power consumption than the existing algorithm, which is greatly beneficial to the hardware implementation of the event camera optical flow prediction., laying the groundwork for future low-power hardware implementation of optical flow prediction for event cameras.

Binary weight network utilizing multi-functional spin-orbit torque device for image recognition

Spintronic devices driven by spin–orbit torque (SOT) have a lot of potential for future neuromorphic computing hardware platforms. Nevertheless, problems such as multistate loss, weight time-dependent variability, and output discontinuity ensue when the device size is reduced to the nanoscale, which are contradictory to the requirements of traditional high-precision neural networks. In this study, we propose a high-precision all-spin neural network based on magnetization switching of a simple nanoscale multi-functional SOT device. The all-spin neural network is implemented by the Binary Weight Network, in which the binary synapse is encoded by the two states formed by the device's deterministic switching and the practical neuron is realized by its stochastic switching probability. Furthermore, we use a difference derivation training algorithm for the general-purpose network in a novel way to be more compatible with the discontinuous neuron output. Using this strategy, our network can reach ∼81.73 % recognition accuracy on the universal CIFAR-10 dataset, paving the way for future practical hardware implementation of nanoscale SOT devices in high-precision compact neuromorphic computing.

Analysis of Statistical and Artificial Intelligence Algorithms for Real-Time Speed Estimation Based on Vehicle Detection with YOLO

Automobiles have increased urban mobility, but traffic accidents have also increased. Therefore, road safety is a significant concern involving academics and government. Transit studies are the main supply for studying road accidents, congestion, and flow traffic, allowing the understanding of traffic flow. They require special equipment (sensors) to measure the car’s speed. With technological advances, artificial intelligence, and videos, it is possible to estimate the speed in real-time without modifying the installed urban infrastructure. We need to employ public databases that provide reliable monocular videos to generate automated traffic studies. The problem of speed estimation with a monocular camera involves synchronizing data recording, tracking, and detecting the vehicles over the road considering the lanes and distance between cars. Usually, a set of constraints are considered, such as camera calibration, flat roads, including methods based on the homography and augmented intrusion lines, patterns or regions, or prior knowledge about the actual dimensions of some of the objects. In this paper, we present a system that generates a dataset from videos recorded from a highway—obtaining 532 samples; we separated the vehicle’s detection by lane, estimating its speed. We use this data set to compare five different statistical methods and three machine learning methods to evaluate their accuracy in estimating the cars’ speed in real-time. Our vehicle estimation requires a feature extraction process using YOLOv3 and Kalman filter to detect and track vehicles. The Linear Regression Model (LRM) yielded the best results obtaining a Mean Absolute Error (MAE) of 1.694 km/h for the center lane and 0.956 km/h for the last lane. The results were compared with several state-of-the-art works, having competitive performance. Hence, LRM is fast estimating speed in real time and does not require high computational resources allowing a future hardware implementation.

“Ghost” and Attention in Binary Neural Network

As the memory footprint requirement and computational scale concerned, the light-weighted Binary Neural Networks (BNNs) have great advantages in limited-resources platforms, such as AIoT (Artificial Intelligence in Internet of Things) edge terminals, wearable and portable devices, etc. However, the binarization process naturally brings considerable information losses and further deteriorates the accuracy. In this article, three aspects are introduced to better the binarized ReActNet accuracy performance with a more low-complex computation. Firstly, an improved Binarized Ghost Module (BGM) for the ReActNet is proposed to increase the feature maps information. At the same time, the computational scale of this structure is still kept at a very low level. Secondly, we propose a new Label-aware Loss Function (LLF) in the penultimate layer as a supervisor which takes the label information into consideration. This auxiliary loss function makes each category’s feature vectors more separate, and improve the final fully-connected layer’s classification accuracy accordingly. Thirdly, the Normalization-based Attention Module (NAM) method is adopted to regulate the activation flow. The module helps to avoid the gradient saturation problem. With these three approaches, our improved binarized network outperforms the other state-of-the-art methods. It can achieve 71.4% Top-1 accuracy on the ImageNet and 86.45% accuracy on the CIFAR-10 respectively. Meanwhile, its computational scale OPs is the least 0.86×10⁸ compared with the other mainstream BNN models. The experimental results prove the effectiveness of our proposals, and the study is very helpful and promising for the future low-power hardware implementations.

Comparative analysis of elementary and edge-preserving spatial filters in noise removal – a comprehensive study

Preserving the edges and information is one of the main purposes of edge-preserving filters. That is, they’re employed to smooth a picture, and minimize halos, phantoms, and edge blur over the edge. They have a nonlinear relationship between one thing and another. Bilateral filters, anisotropic diffusion filters, directed filters, and trilateral filters are all types of example filters. The filter family may be used in a wide range of image processing tasks, such as denoising, video abstraction, demosaicing, optical flow estimation, stereo matching, tone mapping, style transfer, relighting, and others. The paper gives a clear description of edge-preserving filters, from the heat diffusion equation in ancient times to the present, explaining their numerous applications and detailing their numerous uses. Additionally, mathematical analysis is included, as well as efficient and optimized implementations. The focus is on preserving the boundaries, spikes, and canyons, and the information is given clearly and in detail. Finally, it offers a realistic representation of efficient implementation, as well as a comprehensive research scope for future hardware implementation.

A novel adaptive-rear axles steering controller for an 8 × 8 combat vehicle

Multi-axle vehicles are widely used in several applications such as transportation, industrial, and military field, because of its higher reliability in comparison with conventional two axles vehicles. Despite that, there is a paucity of research studies that consider lateral stability enhancement of these vehicles, especially on rough terrain. This simulation-based research study fills this gap and introduces a new adaptive Active Rear Steering (ARS) controller that improves the lateral stability of an 8x8 combat vehicle for rough-terrain operation. The developed controller is designed utilizing the Integral Sliding Mode Control theory (ISMC) based on Gain-Scheduled Linear Quadratic Regulator (GSLQR). Besides, the GSLQR control gains are optimized by a Genetic Algorithm (GA) toolbox using a new synthesized cost function to ensure asymptotic stability. Furthermore, a new Adaptive-ISMC (AISMC) is introduced by using genetic programming to generate control equations that can replace the developed high-dimension GSLQR gains and facilitate future hardware implementation. The developed controller is evaluated by performing a series of simulation-based Double Lane Change (DLC) maneuvers on several rough terrains. The evaluation is conducted for both high friction and slippery surfaces at high and moderate speed, consequently. The results show high fidelity and robustness of the developed controller in comparison with a previously designed optimal LQR controller.

Detection and segmentation of underwater objects from forward-looking sonar based on a modified Mask RCNN

Nowadays, high-frequency forward-looking sonar is an effective device to obtain the main information of underwater objects. Detection and segmentation of underwater objects are also one of the key topics of current research. Deep learning has shown excellent performance in image features extracting and has been extensively used in image object detection and instance segmentation. With the network depth increasing, training accuracy gets saturated and training parameters also increase rapidly. In this paper, a series of residual blocks are used to build a 32-layer feature extraction network and take place of the Resnet50/101 in Mask RCNN, which reduces the training parameters of the network while guaranteeing the detection performance. The parameters of the proposed network are 29% less than Resnet50 and 50.2% less than Resnet101, which is of great significance for future hardware implementation. In addition, Adagrad optimizer is introduced into this research to improve the detection performance of sonar images. Finally, the object detection results of 500 test sonar images show that the mAP is 96.97% that is only 0.18% less than Resnet50 (97.15%) but more than Resnet101 (95.15%).

Load power and energy management system using proteus visual design software

At present, there is a serious energy crisis around the globe which impacts greatly on the power grid. A smart and user friendly energy management system can control the energy consumption. In this paper, load power and energy management system is developed and simulated using Proteus Visual Design software. The load analysis and measurement techniques are developed for single phase and three phase loads and implemented using Arduino Mega 2560 board. User friendly controls are developed using the visual design feature of the software to control the energy consumption. The load power management system is simulated by measuring the power consumed by various residential loads such as lights, fans, air-conditioners, heaters. The industrial loads are simulated by squirrel cage induction motor. The load analysis summary is displayed on the Arduino 2.8 inch TFT display shield in a table format. The simulaion model is created for future hardware implementation and is tested under various practical input conditions. The Proteus Visual Design software is chosen in this resesarch due to its advantages such as professional PCB layout package, availability of nearly 800 microcontrollers in the library package etc.The developed model successfully measured the energy consumption of several loads and assist the user in controlling the energy usage through automation control techniques. The accuracy of the results shows that the technique and the model developed in this research can be used by engineers, students and hobbyists who are working with energy monitoring systems and smart home applications directly. This is an ongoing project where in the next stage, the hardware design of the simulation model will be implemented and tested in real time application.

Bioacoustic Classification of Antillean Manatee Vocalization Spectrograms Using Deep Convolutional Neural Networks

We evaluated the potential of using convolutional neural networks in classifying spectrograms of Antillean manatee (Trichechus manatus manatus) vocalizations. Spectrograms using binary, linear and logarithmic amplitude formats were considered. Two deep convolutional neural networks (DCNN) architectures were tested: linear (fixed filter size) and pyramidal (incremental filter size). Six experiments were devised for testing the accuracy obtained for each spectrogram representation and architecture combination. Results show that binary spectrograms with both linear and pyramidal architectures with dropout provide a classification rate of 94–99% on the training and 92–98% on the testing set, respectively. The pyramidal network presents a shorter training and inference time. Results from the convolutional neural networks (CNN) are substantially better when compared with a signal processing fast Fourier transform (FFT)-based harmonic search approach in terms of accuracy and F1 Score. Taken together, these results prove the validity of using spectrograms and using DCNNs for manatee vocalization classification. These results can be used to improve future software and hardware implementations for the estimation of the manatee population in Panama.

New Sector Scan Geometry for High Frame Rate 2D-Echocardiography using Phased Arrays

2D echocardiography high frame rate techniques do have some drawbacks such as crosstalk artifacts caused by the interactions between the parallel transmitted and received beams. In this paper, we suggest a new cardiac imaging technique based on MLT (Multi-Line Transmission). The main idea of our approach is to benefit from the scan geometry to reduce the interference between the simultaneously transmitted beams. We propose to do the scan at different depths and in parallel to the diagonal scan sector. Therefore, compared to existing MLT techniques, the new scan sector strategy will result in artifacts' reduction in the ultrasound imaging systems. We entitled our approach the Synthetic Sum of Multi-Line Transmission (SS-MLT). Simulations of Point Spread Function (PSF), multiple Point Spread Functions (PSFs) and Cyst Phantom (CP) provided in this paper are compared in our approach to the main classical ultrasound imaging approaches. Therefore, the SS-MLT exhibits a very similar lateral profile as the Single Line Transmission (SLT) algorithm. Hence, the simulation results indicate a potential value of a future hardware implementation of SS-MLT technique.

An Efficient Method of Power Spectral Density Estimation for On-Chip IR-UWB Transmitter Self-Calibration

In ultra-wideband impulse radio communications, process, voltage, and temperature variations may impact the operation of emitters by modifying the emitted power spectral density. In order to compensate the impact of such variations and to respect the spectral emission masks, self-calibration systems able to estimate the output power spectral density have to be integrated in ultra-wideband impulse radio transmitters. In this paper, a new power spectral density estimation method for ultra-wideband impulse radio signals is presented. Unlike the discrete Fourier transform classically used to estimate the power spectral density, the proposed method, which is based on the measurement of the time varying envelope and instantaneous frequency of the emitted pulse, preserves a link between the time and the frequency domain thanks to a piece-wise approximation of these waveforms. This time-to-frequency link enables a straightforward pulse shape control by using simple control methods and allows to comply with spectral masks required by ultra-wideband communication regulations and standards such as the IEEE 802.15.4a. Thus, both envelope and instantaneous frequency extraction techniques are introduced for acquisition of these time domain waveforms in ultra-wideband impulse radio domain. The proposed techniques are validated and allow to consider future hardware implementation in order to build on-chip calibration systems for ultra-wideband transmitters.

Automatic Detection of Epilepsy and Seizure Using Multiclass Sparse Extreme Learning Machine Classification.

An automatic detection system for distinguishing normal, ictal, and interictal electroencephalogram (EEG) signals is of great help in clinical practice. This paper presents a three-class classification system based on discrete wavelet transform (DWT) and the nonlinear sparse extreme learning machine (SELM) for epilepsy and epileptic seizure detection. Three-level lifting DWT using Daubechies order 4 wavelet is introduced to decompose EEG signals into delta, theta, alpha, and beta subbands. Considering classification accuracy and computational complexity, the maximum and standard deviation values of each subband are computed to create an eight-dimensional feature vector. After comparing five multiclass SELM strategies, the one-against-one strategy with the highest accuracy is chosen for the three-class classification system. The performance of the designed three-class classification system is tested with publicly available epilepsy dataset. The results show that the system achieves high enough classification accuracy by combining the SELM and DWT and reduces training and testing time by decreasing computational complexity and feature dimension. With excellent classification performance and low computation complexity, this three-class classification system can be utilized for practical epileptic EEG detection, and it offers great potentials for portable automatic epilepsy and seizure detection system in the future hardware implementation.

Performance and robustness of bio-inspired digital liquid state machines: A case study of speech recognition

This paper presents a systematic performance and robustness study of bio-inspired digital liquid state machines (LSMs) for the purpose of future hardware implementation. Our work focuses not only on the study of the relation between a broad range of network parameters and performance, but also on the impact of process variability and environmental noise on the bio-inspired LSMs from a circuit implementation perspective. In order to shed light on the implementation of LSMs in digital CMOS technologies, we study the trade-offs between hardware overhead (i.e. precision of synaptic weights and membrane voltage and size of the reservoir) and performance. Assisted with theoretical analysis, we leverage the inherent redundancy of the targeted spiking neural networks to achieve both high performance and low hardware cost for the application of speech recognition. In addition, by modeling several types of catastrophic failure and random error, we show that the LSMs are generally robust. Using three subsets of the TI46 speech corpus to benchmark, we elucidate that in terms of isolated word recognition, the analyzed digital LSMs are very promising for future hardware implementation because of their low overhead, good robustness, and high recognition performance.

Embedded processing methods for online visual analysis of laser welding

Online monitoring and closed-loop control of laser welding offer great possibilities for achieving better weld quality. Earlier work on visual laser welding monitoring has mainly focused on aluminum and fairly thin steel used, for example, in car production. We extend this work by focusing on the automated analysis of the phenomena present in the laser welding of thick steel, where all of the phenomena related to the weld quality are still not well understood or controlled. This paper presents the implementation, test results and analysis for weld monitoring methods implemented on a compact smart camera system. The applied embedded sensor–processor platform allows for high-speed implementation of image capture and dynamic range compression, real-time seam tracking and spatter feature extraction. The paper describes experimental results from implemented real-time algorithms for seam tracking and spatter extraction and additional off-line analysis of methods for spatter tracking and seam widening detection, which are also feasible for future online hardware implementation. The results suggest that it is possible to integrate a compact laser welding analysis system, which achieves analysis rates that are sufficient for real-time process control.

Recurrent Spiking Networks Solve Planning Tasks.

A recurrent spiking neural network is proposed that implements planning as probabilistic inference for finite and infinite horizon tasks. The architecture splits this problem into two parts: The stochastic transient firing of the network embodies the dynamics of the planning task. With appropriate injected input this dynamics is shaped to generate high-reward state trajectories. A general class of reward-modulated plasticity rules for these afferent synapses is presented. The updates optimize the likelihood of getting a reward through a variant of an Expectation Maximization algorithm and learning is guaranteed to convergence to a local maximum. We find that the network dynamics are qualitatively similar to transient firing patterns during planning and foraging in the hippocampus of awake behaving rats. The model extends classical attractor models and provides a testable prediction on identifying modulating contextual information. In a real robot arm reaching and obstacle avoidance task the ability to represent multiple task solutions is investigated. The neural planning method with its local update rules provides the basis for future neuromorphic hardware implementations with promising potentials like large data processing abilities and early initiation of strategies to avoid dangerous situations in robot co-worker scenarios.

Hierarchical Adaptive Means (HAM) clustering for hardware-efficient, unsupervised and real-time spike sorting.

This work presents a novel unsupervised algorithm for real-time adaptive clustering of neural spike data (spike sorting). The proposed Hierarchical Adaptive Means (HAM) clustering method combines centroid-based clustering with hierarchical cluster connectivity to classify incoming spikes using groups of clusters. It is described how the proposed method can adaptively track the incoming spike data without requiring any past history, iteration or training and autonomously determines the number of spike classes. Its performance (classification accuracy) has been tested using multiple datasets (both simulated and recorded) achieving a near-identical accuracy compared to k-means (using 10-iterations and provided with the number of spike classes). Also, its robustness in applying to different feature extraction methods has been demonstrated by achieving classification accuracies above 80% across multiple datasets. Last but crucially, its low complexity, that has been quantified through both memory and computation requirements makes this method hugely attractive for future hardware implementation.

A System for the Analysis of Snore Signals

Abstract Sleep apnoea syndrome (SAS) is a disease consisting in the nocturnal cessation of oronasal airflow at least 10 seconds in duration. The standard method for SAS diagnosis is the polysomnographic exam (PSG). However it does not permit a mass screening because it has high cost and requires long term monitoring. This paper presents a preliminary software system prototype for snoring signal analysis, whose main goal is to support the doctor in SAS diagnosis and patient follow-up. The design of the system is modular to allow a future hardware implementation in a portable device for personal snore collection and monitoring.

Optimization of nuclear reactor core fuel reload using the new Quantum PBIL

An issue of great interest in nuclear engineering is to optimize the reload of fuel assemblies in the reactor core, which means to find the best configuration of shuffling between the fresh fuel and the remnants ones from previous cycles. Quantum inspired evolutionary algorithms were developed as an alternative to make the conventional evolutionary algorithms more efficient regarding future hardware implementations. This paper presents a new quantum inspired evolutionary algorithm, named Quantum PBIL (QPBIL). It combines the basic concepts of Population-Based Incremental Learning (PBIL) with the concepts of quantum computing as quantum bit and the linear superposition of states used in evolutionary algorithms with quantum inspirations. To prove its effectiveness as an optimization tool, QPBIL was applied to the optimization of cycle 7 of Angra 1, and the results obtained were comparable to those of efficient optimization techniques based on artificial intelligence currently available.

Towards Real-Time Networked Embedded Generalized Predictive Control for Automotive Active Suspension Systems

This paper emphasizes the implementation challenge of real-time embedded generalized predictive control (GPC) when applied to a class of automotive active suspension systems. A real-time experimental environment is specially designed in order to verify parameters tuning in closed-loop with the virtual process environment. Real-time experimentation carried on this environment shows the necessity of high prediction and control horizons setting to achieve acceptable comfort performance. A control algorithm is decomposed into parallel real-time computations using a set of systolic array based processors. A systemC model is constructed to verify the proposed implementation. Simulation and profiling results show the feasibility and the efficiency of GPC in future hardware implementation.