Backward Path Research Articles

The method to navigate the forward and backward path of a towing tractor for transporting aircraft

Aircraft carriers are the backbone of the Navy. They are equipped with aircraft, and their ability to take off and land aircraft quickly and efficiently determines their performance. Therefore, the number of aircraft is a key consideration in the design of an aircraft carrier and ensuring that they can be operated in the space of the aircraft carrier is an important factor in the arrangement design as well. On the other hands, towing tractors are used to move aircraft around. Towing tractors must safely move aircraft in open spaces while avoiding multiple obstacles, which requires skilled operators. In this study, we propose a method to automate the path of a towing tractor and then follow it. First, we studied the kinematics of the towing tractor and aircraft carrier, considering the wheel movement and steering angle. Then, we calculated the optimal path of the tractor and aircraft considering both forward and backward motion. Finally, we applied dynamics to verify that the towing tractor and aircraft carrier could follow the calculated path. We tested the proposed method in a field with various obstacles and in a narrow area such as a parking lot and confirmed that it was effective.

Supervised Local Training With Backward Links for Deep Neural Networks

The restricted training pattern in the standard BP requires end-to-end error propagation, causing large memory costs and prohibiting model parallelization. Existing local training methods aim to resolve the training obstacles by completely cutting off the backward path between modules and isolating their gradients. These methods prevent information exchange between modules and result in inferior performance. This work proposes a novel local training algorithm, BackLink, which introduces inter-module backward dependency and facilitates information to flow backward along with the network. To preserve the computational advantage of local training, BackLink restricts the error propagation length within the module. Extensive experiments performed in various deep convolutional neural networks demonstrate that our method consistently improves the classification performance of local training algorithms over other methods. For example, our method can surpass the conventional greedy local training method by 6.45% in accuracy in ResNet32 classifying CIFAR100 and recent work by 2.58% in ResNet110 classifying STL-10 with much lower complexity, respectively. Analysis of computational costs reveals that small overheads are incurred in GPU memory costs and runtime on multiple GPUs. Our method can lead up to a 79% reduction in memory cost and 52% in simulation runtime in ResNet110 compared to the standard BP. Therefore, our method could create new opportunities for improving training algorithms towards better efficiency for real-time learning applications.

TargetTree-RRT*: Continuous-Curvature Path Planning Algorithm for Autonomous Parking in Complex Environments

Rapidly-exploring random tree (RRT) has been studied for autonomous parking as it quickly finds an initial path and is easily scalable in complex environments. However, the planning time increases by searching for the path in narrow parking spots. To reduce the planning time, the target tree algorithm, which substitutes a parking goal in RRT with a set (target tree) of backward parking paths, was proposed. However, as it consists of circular and straight paths, it deteriorates parking accuracy because of curvature-discontinuity. Moreover, the planning time increases in complex environments; backward paths can be blocked by obstacles. Therefore, this paper introduces the TargetTree-RRT* algorithm for complex environments. First, a target tree is designed using clothoid paths to address such curvature-discontinuity. Second, to reduce the planning time further, a cost function is defined to initialize a proper target tree that considers obstacles. By integrating with optimal-variant RRT and searching for the shortest path, the proposed TargetTree-RRT* algorithm obtains a near-optimal path as the sampling time increases. Experiment results in real environments showed that the vehicle parked more accurately, and continuous-curvature paths were obtained more quickly and with higher success rates than those acquired using other sampling-based and other types of planning algorithms. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —This work was motivated by the need to develop a fast and practical path planning algorithm for autonomous parking, not only in simple environments but also in complex environments. Sampling-based planning algorithms have been studied in this area, with the advantages of rapid planning of an initial path and easily reflecting vehicle’s constraints. Nevertheless, in cluttered environments, it needs considerable time to obtain a path for the autonomous vehicle easy for tracking and precise parking. To address the aforementioned issue, this article presents the TargetTree-RRT* algorithm. It finds a path by replacing a parking goal with a set of pre-defined continuous-curvature paths considering cluttered environments. TargetTree-RRT* achieves better performance for real parking scenarios and rapidly obtains a near-optimal parking path, compared to other parking path planners. The proposed algorithm is not limited to autonomous vehicles. It can be applied to other unmanned vehicles or robots, such as autonomous underwater vehicles (AUVs). In any application where the goal of path planning can be replaced with predefined standardized paths, the proposed algorithm can be extended.

Adaptive Synaptic Scaling in Spiking Networks for Continual Learning and Enhanced Robustness.

Synaptic plasticity plays a critical role in the expression power of brain neural networks. Among diverse plasticity rules, synaptic scaling presents indispensable effects on homeostasis maintenance and synaptic strength regulation. In the current modeling of brain-inspired spiking neural networks (SNN), backpropagation through time is widely adopted because it can achieve high performance using a small number of time steps. Nevertheless, the synaptic scaling mechanism has not yet been well touched. In this work, we propose an experience-dependent adaptive synaptic scaling mechanism (AS-SNN) for spiking neural networks. The learning process has two stages: First, in the forward path, adaptive short-term potentiation or depression is triggered for each synapse according to afferent stimuli intensity accumulated by presynaptic historical neural activities. Second, in the backward path, long-term consolidation is executed through gradient signals regulated by the corresponding scaling factor. This mechanism shapes the pattern selectivity of synapses and the information transfer they mediate. We theoretically prove that the proposed adaptive synaptic scaling function follows a contraction map and finally converges to an expected fixed point, in accordance with state-of-the-art results in three tasks on perturbation resistance, continual learning, and graph learning. Specifically, for the perturbation resistance and continual learning tasks, our approach improves the accuracy on the N-MNIST benchmark over the baseline by 44% and 25%, respectively. An expected firing rate callback and sparse coding can be observed in graph learning. Extensive experiments on ablation study and cost evaluation evidence the effectiveness and efficiency of our nonparametric adaptive scaling method, which demonstrates the great potential of SNN in continual learning and robust learning.

A 2m-Range 711μW Body Channel Communication Transceiver Featuring Dynamically-Sampling Bias-Free Interface Front End.

Body Channel Communication (BCC) utilizes the body surface as a low-loss signal transmission medium, reducing the power consumption of wireless wearable devices. However, the effective communication range on the human body is limited in the state-of-the-art BCC transceivers, where the signal loss between the body surface and the BCC receiver remains one of the main bottlenecks. To reduce the interface loss, a high input impedance is desired by the BCC receiver, but the DC-biasing circuits decrease the input impedance. In this work, a dynamically-sampling IFE is proposed to eliminate the DC voltage bias, resulting in a 90kΩ high input impedance and a 94dB RF-IF conversion gain to reduce the interface loss in long-range BCC applications. The BCC transceiver chip is fabricated in 55nm CMOS process, taking a die area of 0.123mm2. Measured results show that the chip extends the BCC range to 2m for both the forward and backward paths, where the transmitter and receiver consume 711μW power in total.

Sliding Mode Controller for Autonomous Tractor-Trailer Vehicle Reverse Path Tracking

In the past few years, there has been a growing interest among researchers in developing control systems for autonomous vehicles, specifically for tractor-trailer systems. This newfound interest is driven by the potential benefits of enhancing safety, reducing costs, and addressing labor shortages in the industry. Two industries that could reap the rewards of these systems’ advancements are cargo and agriculture transportation. One of the challenging tasks for the truck trailer vehicle is driving in reverse. Backward path tracking of tractor-trailers is a complex control problem with practical applications. The difficulty in controlling the vehicle arises due to its unstable internal dynamics, coupled nonlinear terms, and the under-actuated nature of the system. There is also a limit to the angle at which the steering can be turned before the risk of a jackknife accident increases significantly. In response to these challenges, this paper introduces a robust sliding mode controller designed for path tracking in reverse-driving tractor-trailer systems. The novelty of our work lies in addressing these challenges, which have not been extensively studied in the past. The proposed controller is analyzed, and its performance is tested and verified using different scenarios. The simulation examples show superior control performance, and we anticipate that this novel controller holds the potential to be widely adopted as a fundamental component in the path-tracking algorithms of autonomous truck trailer systems.

End-to-End Diverse Metasurface Design and Evaluation Using an Invertible Neural Network.

Employing deep learning models to design high-performance metasurfaces has garnered significant attention due to its potential benefits in terms of accuracy and efficiency. A deep learning-based metasurface design framework typically comprises a forward prediction path for predicting optical responses and a backward retrieval path for generating geometrical configurations. In the forward design path, a specific geometrical configuration corresponds to a unique optical response. However, in the inverse design path, a single performance metric can correspond to multiple potential designs. This one-to-many mapping poses a significant challenge for deep learning models and can potentially impede their performance. Although representing the inverse path as a probabilistic distribution is a widely adopted method for tackling this problem, accurately capturing the posterior distribution to encompass all potential solutions remains an ongoing challenge. Furthermore, in most pioneering works, the forward and backward paths are captured using separate models. However, the knowledge acquired from the forward path does not contribute to the training of the backward model. This separation of models adds complexity to the system and can hinder the overall efficiency and effectiveness of the design framework. Here, we utilized an invertible neural network (INN) to simultaneously model both the forward and inverse process. Unlike other frameworks, INN focuses on the forward process and implicitly captures a probabilistic model for the inverse process. Given a specific optical response, the INN enables the recovery of the complete posterior over the parameter space. This capability allows for the generation of novel designs that are not present in the training data. Through the integration of the INN with the angular spectrum method, we have developed an efficient and automated end-to-end metasurface design and evaluation framework. This novel approach eliminates the need for human intervention and significantly speeds up the design process. Utilizing this advanced framework, we have effectively designed high-efficiency metalenses and dual-polarization metasurface holograms. This approach extends beyond dielectric metasurface design, serving as a general method for modeling optical inverse design problems in diverse optical fields.

Deformable lung 4DCT image registration via landmark-driven cycle network.

An automated, accurate, and efficient lung four-dimensional computed tomography (4DCT) image registration method is clinically important to quantify respiratory motion for optimal motion management. The purpose of this work is to develop a weakly supervised deep learning method for 4DCT lung deformable image registration (DIR). The landmark-driven cycle network is proposed as a deep learning platform that performs DIR of individual phase datasets in a simulation 4DCT. This proposed network comprises a generator and a discriminator. The generator accepts moving and target CTs as input and outputs the deformation vector fields (DVFs) to match the two CTs. It is optimized during both forward and backward paths to enhance the bi-directionality of DVF generation. Further, the landmarks are used to weakly supervise the generator network. Landmark-driven loss is used to guide the generator's training. The discriminator then judges the realism of the deformed CT to provide extra DVF regularization. We performed four-fold cross-validation on 10 4DCT datasets from the public DIR-Lab dataset and a hold-out test on our clinic dataset, which included 50 4DCT datasets. The DIR-Lab dataset was used to evaluate the performance of the proposed method against other methods in the literature by calculating the DIR-Lab Target Registration Error (TRE). The proposed method outperformed other deep learning-based methods on the DIR-Lab datasets in terms of TRE. Bi-directional and landmark-driven loss were shown to be effective for obtaining high registration accuracy. The mean and standard deviation of TRE for the DIR-Lab datasets was 1.20±0.72mm and the mean absolute error (MAE) and structural similarity index (SSIM) for our datasets were 32.1±11.6 HU and 0.979±0.011, respectively. The landmark-driven cycle network has been validated and tested for automatic deformable image registration of patients' lung 4DCTs with results comparable to or better than competing methods.

Critical Examination of Distance-Gain-Size (DGS) Diagrams of Ultrasonic NDE with Sound Field Calculations.

Ultrasonic non-destructive evaluation, which has been used widely, can detect and size critical flaws in structures. Advances in sound field calculations can further improve its effectiveness. Two calculation methods were used to characterize the relevant sound fields of an ultrasonic transducer and the results were applied to construct and evaluate Distance-Gain-Size (DGS) diagrams, which are useful in flaw sizing. Two published DGS diagrams were found to be deficient because the backward diffraction path was overly simplified and the third one included an arbitrary procedure. Newly constructed DGS diagrams exhibited transducer size dependence, revealing another deficiency in the existing DGS diagrams. However, the extent of the present calculations must be expanded to provide a catalog of DGS diagrams to cover a wide range of practical needs. Details of the new construction method are presented, incorporating two-way diffraction procedures.

Tracing Knowledge Diffusion Trajectories in Scholarly Bitcoin Research: Co-Word and Main Path Analyses

In the burgeoning field of bitcoin research, a cohesive understanding of how knowledge and insights have evolved over time is lacking. This study aims to address this gap through an exploration of 4123 academic articles pertaining to bitcoin. Utilizing co-word analysis and main path analysis (MPA), it uncovers key themes and seminal works that have substantially influenced the field’s progression. The identified clusters, including safe haven, internet of things (IoT), proof of work (PoW), market efficiency, sentiment analysis, digital currency, and privacy, shed light on the multifaceted discourse surrounding bitcoin. The MPA, incorporating both forward and backward local paths, traces an evolving narrative, starting from an in-depth exploration of bitcoin’s structure, anonymity, and contrasts against traditional financial assets. It tracks the shift in focus to broader market dynamics, volatility, speculative nature, and reactions to economic policy fluctuations. The analysis underscores the transformation of bitcoin research, from its beginnings as a decentralized, privacy-oriented currency to its role in global economics and green financing, revealing a complex narrative of an innovative financial instrument to a multifaceted entity. Implications drawn from this analysis include the need for further research on the potential integration of bitcoin within emerging technologies like AI and cybersecurity, the implications of bitcoin’s interplay with traditional financial systems, and the environmental impacts of bitcoin and blockchain utilization. Overall, the current study not only enhances our understanding of the bitcoin field but also charts its dynamic evolution and stimulates further academic inquiry.

A data-driven approach for plasticity using history surrogates: Theory and application in the context of truss structures

Data-driven methods and algorithms show immense potential for future advancements in the modeling and simulation of complex mechanical systems. However, in order to actually exploit this potential, the methods must be extended to consider inelastic and path-dependent material behavior—a step that appears more complex than in conventional material modeling, where this can be achieved with the help of additional, mostly internal, history variables. The effect of such history variables must thereby be transferred to the data-driven framework. This is achieved in the present paper by defining an appropriate history surrogate as well as a so-called propagator. The history surrogate contains tangible quantities that represent the backward path–uniquely in the ideal case–and enrich the conventional database entries of matching pairs of stresses and strains. The propagator defines the update of the history surrogate from one discrete time step to another. As a major advantage, the newly developed method retains the structures of the data-driven algorithm for elastic material behavior and therefore allows a rather straightforward extension of preexisting program codes. Thus, for instance, heterogeneous problems in which purely elastic and elastoplastic materials are present can be solved without further significant coding effort. Furthermore, several material classes covering different inelastic phenomena such as plasticity with isotropic and kinematic hardening as well as phase transformations in shape memory alloys can be considered in our framework.

Path Planning and Navigation of Miniature Serpentine Robot for Bronchoscopy Application.

The miniature serpentine robot can be applied to NOTES (Natural Orifice Transluminal Endoscopic Surgery). In this paper, a bronchoscopy application is addressed. This paper describes the basic mechanical design and control scheme of this miniature serpentine robotic bronchoscopy. In addition, off-line backward path planning and real-time and in situ forward navigation in this miniature serpentine robot are discussed. The proposed backward-path-planning algorithm utilizes the 3D model of a bronchial tree constructed from the synthetization of medical images such as images from CT (Computed Tomography), MRI (Magnetic Resonance Imaging), or X-ray, to define a series of nodes/events backward from the destination, for example, the lesion, to the original starting point, for example, the oral cavity. Accordingly, forward navigation is designed to make sure this series of nodes/events shall be passed/occur from the origin to the destination. This combination of backward-path planning and forward navigation does not require accurate positioning information of the tip of the miniature serpentine robot, which is where the CMOS bronchoscope is located. Collaboratively, a virtual force is introduced to maintain the tip of the miniature serpentine robot at the center of the bronchi. Results show that this method of path planning and navigation of the miniature serpentine robot for bronchoscopy applications works.

Application of unsupervised deep learning to image segmentation and in-situ contact angle measurements in a CO2-water-rock system

Rock surface wettability is a critical property that regulates multiphase flows in porous media, which can be quantified using the surface contact angle (CA). X-ray micro-computed tomography (μCT) provides an effective approach to in-situ measurements of surface CAs. However, the CA measurement accuracy depends significantly on the quality of CT image segmentation, which is the clustering of CT pixels into separate phases. Inspired by this, we developed a deep learning (DL)-based CA measurement workflow. Motivated by the recent tremendous progress in unsupervised learning techniques and aiming to avoid expensive manual data annotations, an unsupervised DL pipeline for CT image segmentation was proposed and implemented, which includes unsupervised model training and post-processing. The unsupervised model training was driven by a novel loss function constrained with feature similarity and spatial continuity and implemented by iterative forward and backward paths; the former clustered the pixel-wise feature vectors extracted by convolution neural networks, whereas the latter updated the parameters using gradient descent. An over-segmentation strategy was adopted for model training. The post-processing steps based on agglomerative hierarchical clustering (AHC) were implemented to further merge the over-segmented model output to the desired cluster number, which is intended to improve the efficiency of image segmentation. The developed unsupervised DL pipeline was compared with other commonly-used image segmentation methods using pixel-wise and physics-based evaluation metrics on a synthetic raw-image dataset, which had a known ground truth. The unsupervised DL pipeline showed the best performance. Next, the segmented images were input to an automatic CA measurement tool, and the results were validated by comparisons with manual measurements. The CA values from the manual and automatic measurements showed similar distributions and statistical properties. The automatic measurement demonstrated a wider spectrum because of the much larger number of measurement data points. The primary novelty of the unsupervised DL pipeline developed in this study lies in the novel loss function and the over-segmentation strategy associated with AHC post-processing. The workflow has been proven an efficient tool for pore-scale wettability characterization, which has a wide range of applications in fundamental studies of multiphase flows in natural porous media, which have critical implications to geological carbon sequestration, hydrocarbon energy recovery, and contaminant transport in groundwater.

Absence of backward infinite paths for first-passage percolation in arbitrary dimension

In first-passage percolation (FPP), one places nonnegative random variables (weights) (te) on the edges of a graph and studies the induced weighted graph metric. We consider FPP on Zd for d≥2 and analyze the geometric properties of geodesics, which are optimizing paths for the metric. Specifically, we address the question of existence of bigeodesics, which are doubly-infinite paths whose subpaths are geodesics. It is a famous conjecture originating from a question of Furstenberg and most strongly supported for d=2 that, for continuously distributed i.i.d. weights, there a.s. are no bigeodesics. We provide the first progress on this question in general dimensions under no unproven assumptions. Our main result is that geodesic graphs, introduced in a previous paper of two of the authors, constructed in any deterministic direction a.s. do not contain doubly-infinite paths. As a consequence, one can construct random graphs of subsequential limits of point-to-hyperplane geodesics, which contain no bigeodesics. This gives evidence that bigeodesics, if they exist, cannot be constructed in a translation-invariant manner as limits of point-to-hyperplane geodesics.

Correct and Crisp Edge Detection Approach Based on Dense Network

Edge detection is a basic problem in computer vision and image processing. The main purpose of edge detection is to identify points with obvious brightness changes in digital images. At present, there are many good detection methods, but most of them do not consider the correctness and crispness of edges at the same time. In order to address this problem, this paper proposes a method based on a deep convolution neural network. The method is mainly based on a dense network that is then combined with the single network structure of a backward refinement path module. The former can detect and retain the feature information between different layers in an image. The latter makes full use of the extracted information so that the low-level detail features and high-level abstract features can be better integrated in the final output. We tested the method on the BIPED data set. The results show that the correctness and crispness of the edges can be balanced in the detection process, and the ODS, OIS and AP of this method reach 0.888, 0.893 and 0.916, respectively. Compared with the state-of-the-art approaches, the proposed method improves the standard evaluation by 3%-5%, and the convergence speed is also significantly improved.

Bidirectional Optical Carrier Frequency Commonality Method for Optical Fiber Delay Measurement

The transmission delay in an accurate time-transfer system using an optical fiber is calculated by accurately measuring the roundtrip delay of the fiber. The delays in the forward and backward paths of the transmission line must be identical to achieve a highly accurate transmission delay compensation. The effect of chromatic dispersion, which occurs when the optical wavelengths of the forward and backward signals are different, and the effect of backscattering, which occurs in the fiber, are crucial and must be considered for higher accuracy. In this study, we present a fiber transmission delay measurement method that uses optical signals of the same wavelength for the forward and backward paths to overcome these problems. Signal separation can be achieved between the forward and backward signals and between the transmission and backscattered signals by shifting the modulation frequency of a sub-carrier of the optical signal at the remote station. The experimental results demonstrate that a highly accurate time-transfer of more than 1 ns can be expected over a 250 km fiber transmission.

Guidance for Systems-Engineering Processes for Subsea Production Systems

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper OTC 31827, “Systems Engineering of Subsea Production Systems,” by Amedeo Marcotulli, Saipem, and David Wilkinson, Endeavor Management. The paper has not been peer reviewed. Copyright 2022 Offshore Technology Conference. Reproduced by permission. As the complexity and cost of subsea production systems (SPS) has increased during recent decades, the requirement for a more-rigorous, systematic approach toward the engineering of such systems has increased. A systems-engineering (SE) guidance document, specifically written to be applicable to SPS, is being developed by members of API Sub-Committee 17 at the time of writing. The objective of the complete paper is to make potential users of the guidance document aware of the benefits associated with the use of formal SE processes when designing complex facilities. SE Fundamentals Systems engineers exercise a high-level understanding of how system elements work together, visualizing complex systems as a set of subsystems that interact recursively. Recursive relationships are critical in SE and define many hierarchical taxonomies, such as the following: - System components consist of subsystems - Project activities consist of subprojects - Organizations consist of suborganizations - Contracts consist of subcontracts However, an elegance exists in the recursive nature of SE because its processes are self-similar and independent of scale. Its fractal nature allows application of the same process at any level and for any size of project, as shown in Fig. 1. The recursive and iterative nature of SE results in information flows in four directions: - Downward path: Requirement information flows from top to bottom, until the lowest elements are fully defined. - Upward path: Integration information flows from the lowest-level element. - Forward path: Interface information flows across elements that are at the same level, in each level. - Backward path: Feedback information informs the design and realization processes on every iteration for continuous improvement. SE formalizes the interfaces between disciplines, streamlining communications and avoiding rework. SE takes control of the iterations and recursions, thus minimizing waste. Therefore, SE reduces engineering cost while optimizing engineering design. SE requires a mindset more akin to an architect’s duties than to those of an engineer. It requires synthesis more than analysis, involving a top-down view of the system that goes further than an intimate analysis of its parts.

A 2D-graph model-based heuristic approach to visual backtracking security vulnerabilities in physical protection systems

The paper presents a heuristic backward path tracing algorithm in combination with optimization EASI model for effective estimation of time delay remaining after detection TTR. The heuristic backward search algorithm is implemented based on a graph-based model representation of a hypothetical facility for visual path tracing and planning. A comparative analysis between the implementation of non-heuristic and heuristic search for the most vulnerable adversary paths identification in terms of the shortest distance (d), probability of detection (PD), and probability of interruption (PI) on different intrusion detection fields is carried out and in detail discussed. The comparison results show that the global optimal solutions can be obtained using the heuristic backward path tracing algorithm with high search efficiency and accuracy. The knotty problem encountered in unknown path generation under uncertainties for calculating TTR and PI is effectively solved for admissible and consistent heuristic function design.

Spaghetti Tracer: A Framework for Tracing Semiregular Filamentous Densities in 3D Tomograms.

Within cells, cytoskeletal filaments are often arranged into loosely aligned bundles. These fibrous bundles are dense enough to exhibit a certain regularity and mean direction, however, their packing is not sufficient to impose a symmetry between—or specific shape on—individual filaments. This intermediate regularity is computationally difficult to handle because individual filaments have a certain directional freedom, however, the filament densities are not well segmented from each other (especially in the presence of noise, such as in cryo-electron tomography). In this paper, we develop a dynamic programming-based framework, Spaghetti Tracer, to characterizing the structural arrangement of filaments in the challenging 3D maps of subcellular components. Assuming that the tomogram can be rotated such that the filaments are oriented in a mean direction, the proposed framework first identifies local seed points for candidate filament segments, which are then grown from the seeds using a dynamic programming algorithm. We validate various algorithmic variations of our framework on simulated tomograms that closely mimic the noise and appearance of experimental maps. As we know the ground truth in the simulated tomograms, the statistical analysis consisting of precision, recall, and F1 scores allows us to optimize the performance of this new approach. We find that a bipyramidal accumulation scheme for path density is superior to straight-line accumulation. In addition, the multiplication of forward and backward path densities provides for an efficient filter that lifts the filament density above the noise level. Resulting from our tests is a robust method that can be expected to perform well (F1 scores 0.86–0.95) under experimental noise conditions.

EdgePipe: Tailoring Pipeline Parallelism With Deep Neural Networks for Volatile Wireless Edge Devices

As intelligence recently moves to the edge to tackle the problems of privacy, scalability, and network bandwidth in the centralized intelligence, it is necessary to construct an efficient yet robust deep learning model viable at edge devices, which are usually volatile in wireless links and device functionality. The intensive computation burden for deep learning at the edge side necessitates some level of parallel processing via acceleration. We propose <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">EdgePipe</i> , a deep learning framework based on deep neural networks (DNNs) with a mixture of model parallelism and pipeline training for high resource utilization over volatile wireless edge devices. To tackle the volatility problem in wireless links and device functionality, a concept of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">super neuron</i> is defined to be a group of neurons across adjacent layers, which is the basis of model partitioning at edge devices. The relatively loss-resilient neuron structure prevents the entire forward or backward training paths from being totally broken down due to only some intermittent link or device failure caused by one or few devices. Furthermore, we design a subsequent pipeline training mechanism based on the prior super-neuron-based model partitioning for fast convergence with more training data in a fixed timeline. The experimental results have demonstrated that <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">EdgePipe</i> outperforms several counterpart algorithms including <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PipeDream</i> under the volatile wireless lossy or device malfunctioning environments, while preserving the low interlayer communication overhead.