Acceleration Method Research Articles

Compressing convolutional neural networks with cheap convolutions and online distillation

Visual impairment assistance systems play a vital role in improving the standard of living for visually impaired people (VIP). With the development of deep learning technologies and assistive devices, many assistive technologies for VIP have achieved remarkable success in environmental perception and navigation. In particular, convolutional neural network (CNN)-based models have surpassed the level of human recognition and achieved a strong generalization ability. However, the large memory and computation consumption in CNNs have been one of the main barriers to deploying them into resource-limited systems for visual impairment assistance applications. To this end, most cheap convolutions (e.g., group convolution, depth-wise convolution, and shift convolution) have recently been used for memory and computation reduction but with a specific architecture design. Furthermore, it results in a low discriminability of the compressed networks by directly replacing the standard convolution with these cheap ones. In this paper, we propose to use knowledge distillation to improve the performance of compact student networks with cheap convolutions. In our case, the teacher is a network with the standard convolution, while the student is a simple transformation of the teacher architecture without complicated redesigning. In particular, we introduce a novel online distillation method, which online constructs the teacher network without pre-training and conducts mutual learning between the teacher and student network, to improve the performance of the student model. Extensive experiments demonstrate that the proposed approach achieves superior performance to simultaneously reduce memory and computation overhead of cutting-edge CNNs on different datasets, including CIFAR-10/100 and ImageNet ILSVRC 2012, compared to the previous CNN compression and acceleration methods. The codes are publicly available at https://github.com/EthanZhangYC/OD-cheap-convolution.

Characterization of Effects of Compressed Sensing on High Spectral and Spatial Resolution (HiSS) MRI with Comparison to SENSE.

High Spectral and Spatial resolution (HiSS) MRI shows high diagnostic performance in the breast. Acceleration methods based on k-space undersampling could allow stronger T2*-based image contrast and/or higher spectral resolution, potentially increasing diagnostic performance. An agar/oil phantom was prepared with water-fat boundaries perpendicular to the readout and phase encoding directions in a breast coil. HiSS MRI was acquired at 3T, at sensitivity encoding (SENSE) acceleration factors R of up to 10, and the R = 1 dataset was used to simulate corresponding compressed sensing (CS) accelerations. Image quality was evaluated by quantifying noise and artifact levels. Effective spatial resolution was determined via modulation transfer function analysis. Dispersion vs. absorption (DISPA) analysis and full width at half maximum (FWHM) quantified spectral lineshape changes. Noise levels remained constant with R for CS but amplified with SENSE. SENSE preserved the spatial resolution of HiSS MRI, while CS reduced it in the phase encoding direction. SENSE showed no effect on FWHM or DISPA markers, while CS increased FWHM. Thus, CS might perform better in noise-limited or geometrically constrained applications, but in geometric configurations specific to breast MRI, spectral analysis might be compromised, decreasing the diagnostic performance of HiSS MRI.

A diffusion synthetic acceleration method for steady discrete unified gas kinetic scheme in radiative heat transfer

An efficient synthetic-type iterative method is proposed to improve the convergence rate of the steady discrete unified gas kinetic scheme (SDUGKS) for steady radiative heat transfer problems. The key ingredient of the present method is the tight coupling of the mesoscopic radiative transfer equation (RTE) and the macroscopic diffusion equation. The RTE is calculated by SDUGKS, and the diffusion equation is solved to predict the incident radiation energy and angular radiation intensity used in SDUGKS to accelerate the evolution, unlike the original diffusion synthetic acceleration (DSA) method which only predicts the incident radiation energy. Furthermore, a damping factor is adopted to remedy the inconsistency of spatial discretization between the mesoscopic SDUGKS and macroscopic diffusion equation. The theoretical convergence rate of the present method is studied via the discrete Fourier analysis. Several numerical tests under a wide range of optical thicknesses are performed to show the convergence speed. The damping factor and the correction of the angular radiation intensity added in solving the macroscopic diffusion equation greatly increase the stability of the present method and also reduce its computational costs. The convergence rate of the present method becomes faster than that of previous SDUGKS when the optical thickness is larger than 10. With the increase of extinction coefficient, the speedup ratio of the present method and SDUGKS in one-dimensional problems becomes extremely large and also reaches one or two orders of magnitude in two- and three-dimensional problems.

An internal coupling method between neutronics and thermal-hydraulics with RMC and CTF

To tackle the high-fidelity multi-physics coupling challenge in nuclear reactors, an internal coupling method was proposed and implemented to closely couple the neutron transport Monte Carlo code RMC and the sub-channel thermal-hydraulics solver COBRA-TF (CTF) using direct memory data transfers. In this coupling system, fixed-point Picard iteration scheme and Gauss-Seidel iteration method were implemented and evaluated. Two acceleration algorithms were proposed and both proved effective, including the hybrid MPI/OpenMP parallelism scheme and the asymptotic batch coupling method. The coupling system was validated against the VERA benchmarks and compared with literatures. The k eigenvalue differed by 7 pcm and the relative differences in the radial power profiles were less than 0.30%. Moreover, the simulation time was reduced by 48.1% due to quicker convergence with the two acceleration methods. Further analyses also proved the superior precision of this method, especially for asymmetric problems.

Intelligent patrol terminal of transmission line based on AI chip accelerated calculation

With the rapid development of artificial intelligence and information recognition technology, the edge intelligent terminal based on AI chip provides a new way for power patrol. AI chips have become one of the most important components in edge intelligent terminals. However, there are many types of AI chips, and it is particularly important to select appropriate AI chips for application scenarios. Due to the limited computing power of the edge side AI chip, to ensure the real-time performance of the patrol task, it is necessary to research the edge side inference acceleration processing technology. First of all, the parameters and characteristics of various AI chips are analyzed in detail to provide reference for chip selection in different application scenarios. Secondly, a basic framework of power intelligent patrol terminal based on AI chip is proposed, which includes hardware component and software component. Then, in order to transplant the neural network model on the server side to the intelligent patrol terminal, the edge side inference acceleration processing technology is analyzed in detail, specifically the neural network acceleration and compression method and the inference acceleration technology based on TensorRT. Finally, in order to prove the feasibility and effectiveness of the whole scheme, an intelligent patrol terminal applied to the visual defect recognition for transmission lines has been developed and tested.

An evaluation of Kd-Trees vs Bounding Volume Hierarchy (BVH) acceleration structures in modern CPU architectures

Ray tracing is a rendering technique that is highly praised for its realism and image quality. Nonetheless, this is a computationally intensive task that is slow compared to other rendering techniques like rasterization. Bounding Volume Hierarchy (BVH) is a primitive subdivision acceleration mechanism that is the mainly used method for accelerating ray tracing in modern solutions. It is regarded as having better performance against other acceleration methods. Another well-known technique is Kd-Trees that uses binary space partitioning to adaptively subdivide space with planes. In this research, we made an up-to-date evaluation of both acceleration structures, using state-of-the-art BVH and Kd-Trees algorithms implemented in C, and found out that the Kd-Trees acceleration structure provided better performance in all defined scenarios on a modern x86 CPU architecture.

ANALISIS PENJADWALAN DENGAN PERCEPATAN PELAKSANAAN MENGGUNAKAN METODE PENAMBAHAN JUMLAH TENAGA KERJA DAN JAM KERJA

This study aims to analyze the factors that cause delays and analyze the time and cost of the acceleration method with the addition of the number of workers and the addition of working hours to obtain a more efficient comparison of the acceleration method for project completion. As for the steps in this study, namely the collection of data in the form of questionnaire results and existing project data, including the Budget Plan (RAB), Analysis of Work Unit Prices (AHSP) and Time Schedule. Factors causing project delays based on questionnaires were analyzed using interest index calculations with validity and reliability testing, in order to obtain the top ranking of factors causing delays with a value of 𝑟ℎ = 0.502 < 𝑟𝑡 = 0.576 with "valid" test results caused by changes in material or material specifications. The results of the analysis of project plan conditions with a duration of 369 days, direct costs of Rp. 11,092,954,636.14 and an indirect cost of Rp. 2,160,276,709.31, and the results of the acceleration analysis with the addition of the number of workers (alternative 1) obtained a duration of 350 days with a direct cost of Rp. 11,092,954,636.14 and an indirect cost of Rp. 2,160,276,709.31, while the results of the acceleration analysis with the addition of working hours (alternative 2) obtained a duration of 364 days with a direct cost of Rp. 11,648,582,543.10 and an indirect cost of Rp. 1,604,648,802.35.

Assessment of Dentoalveolar Changes Following Leveling and Alignment of Severely Crowded Upper Anterior Teeth Using Self-Ligating Brackets Alone or With Flapless Piezocision Compared to Traditional Brackets: A Randomized Controlled Clinical Trial.

Introduction Dental crowding is one of the most common types of malocclusions. It can be treated with or without extraction, depending on the severity of the crowding. Extraction-based orthodontic treatments are the preferred treatment option in cases of severe crowding, but they take longer than non-extraction cases. Objective This study aimed to evaluate the dentoalveolar changes following the orthodontic treatment of severely crowded maxillary anterior teeth in adults using self-ligating brackets alone or combined with flapless piezocision. Materials and methods The participants in this study were 63 patients (46 females and 17 males; mean age SD: 19.71 ± 2.74 years) who attended the Department of Orthodontics at the University of Damascus from January 2020 to December 2021. The participants were divided into three groups at random: Group (1): traditional brackets group, Group (2): self-ligating brackets group; and Group (3): self-ligating brackets with flapless piezocision group. Little's Irregularity Index (LII) was measured at five assessment times: before the onset of orthodontic treatment (T0), after one month (T1), after two months (T2), after three months (T3), and at the end of the leveling and alignment phase (T4). The intercanine width (lingual), the intercanine width (cusp), and the canine rotation angle were measured at two assessment times: before the onset of orthodontic treatment (T0) and at the end of the leveling and alignment phase (T4). Results The three studied groups had statistically significant differences in terms of LII during the first three months, and the most significant improvement of LII was in the self-ligating brackets with the piezocision group (P < 0.001). In addition, the intercanine width (cusp) at the end of the leveling and aligning phase revealed greater mean values in both self-ligating brackets groups compared to the traditional brackets group, and the differences were statistically significant (P < 0.001). Otherwise, no statistically significant differences were found at the end of the leveling and aligning phase in the intercanine width (lingual) or the canine rotation angle between the three studied groups (P > 0.05). Conclusion Using self-ligating brackets with flapless piezocision revealed more significant results concerning LII as compared to other groups. Thus, combining these two acceleration methods could get more effective results in aligning severely crowded teeth. Self-ligating brackets, whether used alone or with flapless piezocision, resulted in greater intercanine width at the cusplevel. The type of brackets (traditional or self-ligating) did not affect the canine rotation angle.

STABILITY ANALYSIS OF THE SLOPE SUBJECTED TO THE DYNAMIC LOADING

This paper presents the results of a stability analysis of a slope located in the immediate vicinity of a railway line. The plans for the extension of this railway track include the construction of another line, which would run parallel to the existing one, within a few metres distance. It is expected that intensive goods train traffic will generate both static and dynamic forces in the underlying subsoil. Consequently, seismic vibrations will be generated in the subsoil, propagating mainly not only in the horizontal direction but also in the vertical direction. The method of seismic coefficient of the earthquake intensity determined by a pseudo-static method and horizontal component of acceleration is appropriate and recommended because it is simple, and the safety factor of the slope is calculated in the same way as in conventional stability calculations.

High-Resolution 3D Magnetic Resonance Fingerprinting With a Graph Convolutional Network.

Magnetic resonance fingerprinting (MRF) is a novel quantitative imaging framework for rapid and simultaneous quantification of multiple tissue properties. 3D MRF allows higher through-plane resolution, but the acquisition process is slow when whole-brain coverage is needed. Existing methods for acceleration mainly rely on GRAPPA for k-space interpolation in the partition-encoding direction, limiting the acceleration factor to 2 or 3. In this work, we replace GRAPPA with a deep learning approach for accurate tissue quantification with greater acceleration. Specifically, a graph convolution network (GCN) is developed to cater to the non-Cartesian spiral sampling trajectories typical in MRF acquisition. The GCN maintains high quantification accuracy with up to 6-fold acceleration and allows 1mm isotropic resolution whole-brain 3D MRF data to be acquired in 3min and submillimeter 3D MRF (0.8mm) in 5min, greatly improving the feasibility of MRF in clinical settings.

The Acceleration Phenomenon of Shock Wave Induced by Nanosecond Laser Irradiating Silicon Assisted by Millisecond Laser

The propagating evolution of shock waves induced by a nanosecond pulse laser (ns laser) irradiating silicon assisted by a millisecond pulse laser (ms laser) is investigated experimentally. A numerical model of 2D axisymmetric two-phase flow is established to obtain the spatial distribution of shock wave velocity. Two types of shock wave acceleration phenomenon are found. The mechanism of the shock wave acceleration phenomenon is discussed. The experimental and numerical results show that the initial stage of ms laser-induced plasma can provide the initial ions to increase probability of collision ionization between free electrons and vapor atoms. The velocity of the ns laser-induced shock wave is accelerated. Furthermore, the ms laser-induced plasma as the propagation medium can also accelerate the ns laser-induced shock wave. The shock wave acceleration methods obtained in this paper can promote the development of laser propulsion technology.

Fast density estimation for density-based clustering methods

Density-based clustering algorithms are widely used for discovering clusters in pattern recognition and machine learning. They can deal with non-hyperspherical clusters and are robust to outliers. However, the runtime of density-based algorithms is heavily dominated by neighborhood finding and density estimation which is time-consuming. Meanwhile, the traditional acceleration methods using indexing techniques such as KD-tree may not be effective when the dimension of the data increases. To address these issues, this paper proposes a fast range query algorithm, called Fast Principal Component Analysis Pruning (FPCAP), with the help of the fast principal component analysis technique in conjunction with geometric information provided by the principal attributes of the data. Based on FPCAP, a framework for accelerating density-based clustering algorithms is developed and successfully applied to accelerate the Density Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm and the BLOCK-DBSCAN algorithm, and improved DBSCAN (called IDBSCAN) and improved BLOCK-DBSCAN (called BLOCK-IDBSCAN) are then obtained, respectively. IDBSCAN and BLOCK-IDBSCAN preserve the advantage of DBSCAN and BLOCK-DBSCAN, respectively, while greatly reducing the computation of redundant distances. Experiments on seven benchmark datasets demonstrate that the proposed algorithm improves the computational efficiency significantly.

An Efficient High-to-Low Iterative Method for Light Water Reactor Analysis Based on NEAMS Tools

The so-called two-step method involving the consecutive lattice physics and core simulation has been successfully and widely used in large-scale nuclear reactor calculations thanks to its superior computational efficiency and a satisfactory level of accuracy. However, its performance is challenged by the ever-increasing level of heterogeneity in core designs due to the use of infinite lattice approximation in the lattice calculation and its inability to update cross-section sets according to the core environment change. This paper introduces an alternative approach for light water reactor steady-state core analysis. During the core calculation process, iterations between the local lattice transport calculation and the global core nodal calculation are conducted. These iterations continuously update the boundary condition applied to the lattice model and generate updated cross-section sets. This is done through the iteration between the local lattice transport calculation and the global core nodal simulation. The neutronics high-to-low (Hi2Lo) scheme was formulated using Nuclear Energy Advanced Modeling and Simulation or NEAMS codes, in particular, with the modified PROTEUS-MOC and PROTEUS-NODAL serving as the transport lattice solver and full-core nodal solver, respectively. The verification of the implemented Hi2Lo iterative scheme on the two-dimensional C5G7-TD benchmark problem shows that the Hi2Lo scheme outperforms the two-step approach in terms of prediction accuracy for the key responses of interest (e.g., the system eigenvalue and power distribution) at a computational cost lower than that of the direct full-core transport calculation. To further improve its efficiency, an acceleration method has been developed and implemented for the Hi2Lo approach, and the results indicate that the acceleration method can significantly reduce the run time of a full-core transport solution by a factor of 14 while generating solutions with comparable accuracy.

A time-domain method for free vibration responses of an equivalent viscous damped system based on a complex damping model

The free vibration responses are divergent based on a complex time-domain damping model. The traditional step by step integration method cannot be used to calculate the time-domain responses. Based on the theoretic solution that eliminates the divergent term for the complex damping model, the equivalent viscous damping model is proposed and the corresponding average acceleration method is realized in this paper. The numerical cases show that the calculated results of the equivalent viscous damped system are convergent, which are equal to the ones of the deleted divergent term based on the complex damping model. The correctness of the proposed method is verified, and the computational efficiency is high.

Research on seismic analysis methods of large and complex underground pipe structures in hard rock sites

In order to study the seismic design methods of large and complex underground pipe structures (LCUP) in hard rock sites, firstly, the structural characteristics of LCUP and the calculation principles of four typical seismic analysis methods were carefully sorted out. Secondly, based on different seismic analysis methods, the deformation and internal force response of three representative pipes were studied in detail, and the calculation accuracy and error causes of each seismic analysis method were analyzed. Finally, the advantages, disadvantages and applicability of the seismic analysis methods on typical pipes were systematically explained, and the technical route for seismic design of LCUP was proposed. The results showed that both the reaction displacement method and the reaction acceleration method can effectively analyze the seismic performance of standard section structures. The three-dimensional time-history analysis method was comprehensive and highly accurate for seismic analysis of complex spatial cross-node structures. The rigid connections of different pipes leaded to serious discontinuities in stiffness at the intersection, and the deformation and bending moment distributions were highly concentrated at the height away from the intersection. The influence range of the intersection was about three times the width of the intersection, and LCUP was classified according to this influence range. For standard section structures, seismic analysis should be carried out using the reaction acceleration method or reaction displacement method, while for spatial cross-node structures, three-dimensional time-history analysis should be used for seismic analysis. This scientific seismic design method achieved multiple optimal balances of seismic safety, analysis accuracy, efficiency and economy in the seismic analysis of large and complex underground structures. A theoretical basis and useful reference for the scientific design, accurate prediction and seismic safety evaluation of seismic response of LCUP were provided in this paper.

ON GUILLERA’S -SERIES FOR

AbstractIn 2011, Guillera [‘A new Ramanujan-like series for $1/\pi ^2$ ’, Ramanujan J.26 (2011), 369–374] introduced a remarkable rational ${}_{7}F_{6}( \frac {27}{64} )$ -series for ${1}/{\pi ^2}$ using the Wilf–Zeilberger (WZ) method, and Chu and Zhang later proved this evaluation using an acceleration method based on Dougall’s ${}_{5}F_{4}$ -sum. Another proof of Guillera’s ${}_{7}F_{6}( \frac {27}{64} )$ -series was given by Guillera in 2018, and this subsequent proof used a recursive argument involving Dougall’s sum together with the WZ method. Subsequently, Chen and Chu introduced a q-analogue of Guillera’s ${}_{7}F_{6}( \frac {27}{64} )$ -series. The many past research articles concerning Guillera’s ${}_{7}F_{6}( \frac {27}{64} )$ -series for ${1}/{\pi ^2}$ naturally lead to questions about similar results for other mathematical constants. We apply a WZ-based acceleration method to prove new rational ${}_{7}F_{6}( \frac {27}{64} )$ - and ${}_{6}F_{5}( \frac {27}{64} )$ -series for $\sqrt {2}$ .

C-S-H Seeds Accelerate Early Age Hydration of Carbonate-Activated Slag and the Underlying Mechanism

The slow hardening process of carbonate-activated slag limits its application as a construction material. This paper aims to provide an acceleration method for the early age hydration of carbonate-activated slag by applying calcium silicate hydrate (C-S-H) seeds and unveil the underlying mechanism. The results show that the incorporation of C-S-H seeds significantly accelerates the early age reaction of carbonate-activated slag and shortens the setting time. With 4% of calcium silicate hydrate (C-S-H) seeds, the 1d-compressive strength of carbonate activates slag can achieve 25.4 MPa. The C-S-H seeds acts as the preferred nucleation sites for the strength-giving phase C-A-S-H gel and the carbonate-containing phases (e.g., calcite, gaylussite, hydrotalcite, etc.), and accelerates hydration. The dormant period of samples with C-S-H seeds becomes negligible, confirming that the seeding effect that controls the saturation limits of the pore solution is the major reason for the accelerated hydration.

Analysis of The Acceleration of Time and Cost of Implementing Building Construction Projects Using The Critical Path Method (CPM) Method

This study aims to obtain the form of project work network in the Probolinggo City Hospital Development Project, to determine the effect of accelerating the duration of Probolinggo City Hospital Development work on network planning and to find out the amount of costs incurred before and after the acceleration of the duration of project work. The method used is the acceleration method with modification of the scheduling of the CPM model. CPM method planning is a recommendation in choosing from many existing methods, because there is a calculation of the earliest time the project starts, the last time the project starts as well as the earliest time to finish and the last time to finish so that it can be known the time lag of all activities. Data Used are primary and secondary data related to the sequence and relationship of activities, implementation time and cost budget plan (RAB). In this thesis, the author takes scheduling data from the Probolinggo City Hospital Development Project which will be used as the basis for writing this thesis in conducting CPM analysis and accelerating the duration of work with additional overtime working hours and calculating the total cost of the acceleration to be used as comparison material in the writing results. Based on the analysis that has been carried out using the Microsoft Project 2016 application, the normal duration of structural work is 81 working days and has 18 critical units of 31 work items. After the acceleration was carried out, there was a decrease in the duration of the work of the structure to 74 working days. The difference in duration decrease was 7 days or 9.45%. The total initial cost with the normal duration for structural work is IDR 28,563,131,560, while after acceleration, the project cost for structural work becomes IDR 28,635,083,793

A Reactive Molecular Dynamics Study on Crosslinked Epoxy Resin Decomposition under High Electric Field and Thermal Aging Conditions.

To reveal the microscopic mechanism of synergetic thermal-electrical degradation during a partial discharge process in epoxy insulation materials, the decomposition of crosslinked epoxy resin is investigated using reactive molecular dynamics simulations under high electric field and thermal degradation conditions. Bond-boost acceleration method is employed in reactive molecular dynamics simulations to successfully establish epoxy polymer models with a crosslink degree of 93%. Active molecular species derived from electrical partial discharges are considered in the current work. Small molecule products and decomposition temperature in the degradation process under an electric field are calculated to elucidate the effect of nitric acid and ozone molecules, being the active products generated by electrical partial discharges, on the synergetic thermal-electrical degradation of epoxy resin. Both nitric acid and ozone exacerbate thermal impact decomposition of crosslinked epoxy polymer by decreasing initial decomposition temperature from 1050 K to 940 K and 820 K, respectively. It is found that these active products can oxidize hydroxyl groups and carbon-nitrogen bridge bonds in epoxy molecular chains, leading to the aggravation of epoxy resin decomposition, as manifested by the significant increase in the decomposed molecular products. In contrast, thermal degradation of the epoxy resin without the active species is not expedited by increasing electric field. These strongly oxidative molecules are easily reduced to negative ions and able to obtain kinetic energies from electric field, which result in chemical corrosion and local temperature increase to accelerate decomposition of epoxy insulation materials.

An Annealing Mechanism for Adversarial Training Acceleration.

Despite the empirical success in various domains, it has been revealed that deep neural networks are vulnerable to maliciously perturbed input data that can dramatically degrade their performance. These are known as adversarial attacks. To counter adversarial attacks, adversarial training formulated as a form of robust optimization has been demonstrated to be effective. However, conducting adversarial training brings much computational overhead compared with standard training. In order to reduce the computational cost, we propose an annealing mechanism, annealing mechanism for adversarial training acceleration (Amata), to reduce the overhead associated with adversarial training. The proposed Amata is provably convergent, well-motivated from the lens of optimal control theory, and can be combined with existing acceleration methods to further enhance performance. It is demonstrated that, on standard datasets, Amata can achieve similar or better robustness with around 1/3-1/2 the computational time compared with traditional methods. In addition, Amata can be incorporated into other adversarial training acceleration algorithms (e.g., YOPO, Free, Fast, and ATTA), which leads to a further reduction in computational time on large-scale problems.