Low Memory Requirements Research Articles

Efficient Hardware for Generalized Turbo Signal Recovery in Compressed Sensing

Compressed sensing (CS) is becoming a hot topic in recent years for its advantages such as low-power consumption, low memory requirement, and low sampling frequency. However, high-dimensional nonlinear signals will inevitably introduce notable complexity and low efficiency in signal recovery. Generalized turbo signal recovery (G-Turbo-SR) is a cutting-edge method, which efficiently reduces complexity with a partial discrete Fourier transform (DFT) sensing matrix. However, in practical applications, G-Turbo-SR still suffers from high complexity for probability computations and matrix multiplications. This article optimizes the algorithm of G-Turbo-SR in scheduling to reduce the matrix multiplications by half. High-precision numerical approximation method is proposed to replace the complex integral calculation, which efficiently reduces the hardware cost with acceptable performance degradation. Based on the data-flow graph (DFG) analysis, detailed hardware architecture is proposed with module designs. Proper quantization scheme is selected according to the mean square error (MSE) performance. Pipelining, folding, and variable precision quantization (VPQ) scheme are employed for higher hardware efficiency. FPGA implementation on Xilinx 7k325tffg900-2 shows a higher throughput and hardware efficiency compared to existing recovery methods for CS.

Negative Reactance Impacts Power Flow Convergence Using Conjugate Gradient Method

It is usually considered in power systems that the B matrices in fast decoupled power flow (B’ and B”) are symmetric and positive-definite. The fast decoupled power flow (FDPF) based on the conjugate gradient (CG) iterative method was well developed, because the CG iterative method has a good convergence property and a lower memory requirement which can be easily implemented in the parallel computation. However, a rare yet important phenomenon hasn’t been well addressed in that “negative reactance” may exist in the practical power system models, which could affect the definiteness of B matrices in FDPF and the CG convergence performance. In this brief, the eigenvalues of B matrices with negative reactance are investigated and the convergence of CG for FDPF is discussed. Several large-scale practical systems are tested, which could provide some insights for the study of the FDPF using the CG method with negative reactances.

Distilling Channels for Efficient Deep Tracking.

Deep trackers have proven success in visual tracking. Typically, these trackers employ optimally pre-trained deep networks to represent all diverse objects with multi-channel features from some fixed layers. The deep networks employed are usually trained to extract rich knowledge from massive data used in object classification and so they are capable to represent generic objects very well. However, these networks are too complex to represent a specific moving object, leading to poor generalization as well as high computational and memory costs. This paper presents a novel and general framework termed channel distillation to facilitate deep trackers. To validate the effectiveness of channel distillation, we take discriminative correlation filter (DCF) and ECO for example. We demonstrate that an integrated formulation can turn feature compression, response map generation, and model update into a unified energy minimization problem to adaptively select informative feature channels that improve the efficacy of tracking moving objects on the fly. Channel distillation can accurately extract good channels, alleviating the influence of noisy channels and generally reducing the number of channels, as well as adaptively generalizing to different channels and networks. The resulting deep tracker is accurate, fast, and has low memory requirements. Extensive experimental evaluations on popular benchmarks clearly demonstrate the effectiveness and generalizability of our framework.

A novel motion recovery using temporal and spatial correlation for a fast temporal error concealment over H.264 video sequences

In this paper, we proposed an effective motion recovery method with low complexity for a fast temporal error concealment over H.264 video sequences. The proposed algorithm uses temporal and spatial motion vectors of a lost block and neighboring blocks. A neighboring block with minimum SAD-MV between its trajectory and trajectory of lost block is chosen as the most reliable candidate. In order to avoid the mosaic artifacts in the process of recombination of optimal blocks, final motion vector of the block is recalculated using H.264 partition information and then H,.264 1/4 interpolation is followed. The proposed method is faster and better in PSNR compared with previous algorithms. The proposed algorithm may be used as a powerful error concealment tools for real time video applications, such as mobile phone, tactile display and WSN cameras, using H.264/AVC decoder, especially in the ROI tracking applications, because of low memory requirement.

In-cylinder pressure-based convolutional neural network for real-time estimation of low-pressure cooled exhaust gas recirculation in turbocharged gasoline direct injection engines

Low-pressure cooled exhaust gas recirculation is one of the most promising technologies for improving fuel efficiency of turbocharged gasoline direct injection engines. To realize the beneficial effects of the low-pressure cooled exhaust gas recirculation, the accurate estimation of the low-pressure cooled exhaust gas recirculation rate is essential for precise low-pressure cooled exhaust gas recirculation control. In this respect, previous studies have suggested in-cylinder pressure-based low-pressure cooled exhaust gas recirculation models to obtain the low-pressure cooled exhaust gas recirculation rate into the cylinders with fast response. However, these methods require considerable manual process of feature engineering to extract and analyze the combustion characteristics from the cylinder pressure traces. Furthermore, the performance of the entire model is limited solely to certain hand-crafted characteristics and their mathematical formulations. To resolve these limitations, we propose an in-cylinder pressure-based convolutional neural network for low-pressure cooled exhaust gas recirculation estimation. Because the convolutional neural network model automatically learns the complex function between the raw input of the high-dimensional cylinder pressure traces and the low-pressure cooled exhaust gas recirculation rate through an end-to-end deep learning framework, this convolutional neural network model provides a more effective and precise modeling process compared to the conventional combustion characteristics-based regression models. The proposed convolutional neural network model consists of the input layer with the previous consecutive cycles of the pressure traces to resolve the model uncertainty from cycle-to-cycle variations. This input layer is connected to one convolutional layer, two fully connected layers, and the final output layer that is the target low-pressure cooled exhaust gas recirculation rate. The proposed model was trained, validated, and tested using a total of 50,000 cycles of engine experimental data under various transient driving conditions. The remarkable accuracy of the proposed model was evaluated with R2 values over 0.99 and root mean square error values of less than 1.5% under the transient conditions. Moreover, the real-time performance and low memory requirement were also verified on the target embedded platform.

New class of hybrid conjugate gradient coefficients with guaranteed descent and efficient line search

Hybrid conjugate gradient (CG) techniques are one of the most prominent procedure for obtaining the solution of large-scale unconstrained optimization problems. This is due to its simplicity, global convergence, and low memory requirement. Numerous modifications have been done recently to improve the performance of these methods. In this paper, we proposed new class of hybrid CG coefficients with guaranteed descent under exact line search. Numerical results are presented to illustrate the efficiency of the proposed methodscompared to other classical CG coefficients.

Transductive non-linear semantic embedding for multi-class classification

This paper presents a method for semi-supervised multi-class classification based on matrix factorization. The method called semi-supervised online kernel matrix factorization (SS-OKMF) performs a semantic embedding by finding a non-linear mapping to a low-dimensional semantic space. An important characteristic of the SS-OKMF method is that the new low-dimensional semantic representation can be learned in a semi-supervised fashion. Therefore, the annotated instances can be used to maximize the discrimination between classes, but also, the non-annotated instances can be exploited to estimate the intrinsic manifold structure of the data. The non-linear modeling is based on kernel methods with a learning-in-a-budget strategy that allows keeping low memory requirements. This strategy, along with an online formulation based on stochastic gradient descent, reduces the computation time and keeps low computational requirements in large-scale problems. According to the experimental evaluation performed on several datasets of different nature (i.e. images, biosignals, and synthetic data), SS-OKMF, in comparison with several non-linear supervised, unsupervised and semi-supervised dimensionality reduction methods, presents a competitive performance in classification tasks under a transductive learning setup preserving a lower computational cost.

Utilization of the Brinkman Penalization to Represent Geometries in a High-Order Discontinuous Galerkin Scheme on Octree Meshes

We investigate the suitability of the Brinkman penalization method in the context of a high-order discontinuous Galerkin scheme to represent wall boundaries in compressible flow simulations. To evaluate the accuracy of the wall model in the numerical scheme, we use setups with symmetric reflections at the wall. High-order approximations are attractive as they require few degrees of freedom to represent smooth solutions. Low memory requirements are an essential property on modern computing systems with limited memory bandwidth and capability. The high-order discretization is especially useful to represent long traveling waves, due to their small dissipation and dispersion errors. An application where this is important is the direct simulation of aeroacoustic phenomena arising from the fluid motion around obstacles. A significant problem for high-order methods is the proper definition of wall boundary conditions. The description of surfaces needs to match the discretization scheme. One option to achieve a high-order boundary description is to deform elements at the boundary into curved elements. However, creating such curved elements is delicate and prone to numerical instabilities. Immersed boundaries offer an alternative that does not require a modification of the mesh. The Brinkman penalization is such a scheme that allows us to maintain cubical elements and thereby the utilization of efficient numerical algorithms exploiting symmetry properties of the multi-dimensional basis functions. We explain the Brinkman penalization method and its application in our open-source implementation of the discontinuous Galerkin scheme, Ateles. The core of this presentation is the investigation of various penalization parameters. While we investigate the fundamental properties with one-dimensional setups, a two-dimensional reflection of an acoustic pulse at a cylinder shows how the presented method can accurately represent curved walls and maintains the symmetry of the resulting wave patterns.

An Efficient FIR Filter Architecture Implementation using Distributed Arithmetic (DA) for DSP Applications

In this paper the proposed efficient FIR filter architecture using a distributed arithmetic (DA) algorithm in which two issues are discussed in the conventional FIR filter. The FIR filter is well known to include delay elements, multipliers and adders. Due to the need for multipliers, this results in 2 demerits which are (i) increased in area and (ii) delayed increases that eventually lead to low efficiency (low speed). A notable feature of the proposed technique is to substitute a trivial amount of indexed LUT pages instead of conventional LUT based DA that it helps to maintain the access time lower. Also, significant idea connected with the proposed technique is required page can be thoroughly selected with the selection module without needing adders that result in reduced computation time. Furthermore, the proposed fast FIR filter is used for the powerful ECG noise elimination technique, which is prevalently used in biomedical and healthcare applications. The designs are simulated and synthesized by using Xilinx ISE. It can be seen from reports that our proposed DA consumes 30% less power for 11-tap FIR filters with a 40% shorter area, while the saving in power consumption for 8-tap FIR filters is 30% to 80% and 35% to 80% in the area. Especially in contrast with all the above-mentioned DA techniques, our enhanced quick FIR filters require less area and less power intake due to their lower memory requirements. All architectures are designed for FIR filters with 4 and 8 taps.

An efficient hybrid conjugate gradient-based projection method for convex constrained nonlinear monotone equations

Conjugate gradient projection methods are widely used for solving large-scale nonlinear systems of monotone equations because of their low memory requirements and research continues to be done to improve these methods. In this paper, we introduce a new hybrid conjugate gradient projection method for solving large-scale nonlinear monotone equations. This proposed method is shown to satisfy the descent property and its global convergence is also established. Numerical results show that the new method is very much promising when compared with other conjugate gradient projection methods.

Automatic Brain Tumor Segmentation Based on Cascaded Convolutional Neural Networks With Uncertainty Estimation.

Automatic segmentation of brain tumors from medical images is important for clinical assessment and treatment planning of brain tumors. Recent years have seen an increasing use of convolutional neural networks (CNNs) for this task, but most of them use either 2D networks with relatively low memory requirement while ignoring 3D context, or 3D networks exploiting 3D features while with large memory consumption. In addition, existing methods rarely provide uncertainty information associated with the segmentation result. We propose a cascade of CNNs to segment brain tumors with hierarchical subregions from multi-modal Magnetic Resonance images (MRI), and introduce a 2.5D network that is a trade-off between memory consumption, model complexity and receptive field. In addition, we employ test-time augmentation to achieve improved segmentation accuracy, which also provides voxel-wise and structure-wise uncertainty information of the segmentation result. Experiments with BraTS 2017 dataset showed that our cascaded framework with 2.5D CNNs was one of the top performing methods (second-rank) for the BraTS challenge. We also validated our method with BraTS 2018 dataset and found that test-time augmentation improves brain tumor segmentation accuracy and that the resulting uncertainty information can indicate potential mis-segmentations and help to improve segmentation accuracy.

An Extension of Polak-Ribière-Polyak Method using Exact Line Search

Lately, many large-scale unconstrained optimization problems rely upon nonlinear conjugate gradient (CG) methods. Many areas such as engineering, and computer science have benefited because of its simplicity, fast and low memory requirements. Many modified coefficients have appeared recently, all of which to improve these methods. This paper considers an extension conjugate gradient method of PolakRibière-Polyak using exact line search to show that it holds for some properties such as sufficient descent and global convergence. A set of 113 test problems is used to evaluate the performance of the proposed method and get compared to other existing methods using the same line search.

An Efficient Extended Targets Detection Framework Based on Sampling and Spatio-Temporal Detection.

Excellent performance, real-time and low memory requirement are three vital requirements for target detection in high resolution marine radar system. Unfortunately, many current state-of-the-art methods merely achieve excellent performance when coping with highly complex scenes. In fact, a common problem is that real-time processing, low memory requirement and remarkable detection ability are difficult to coordinate. To address this issue, we propose a novel detection framework which bases its principle on sampling and spatiotemporal detection. The framework consists of two stages, coarse detection and fine detection. Sampling-based coarse detection is designed to guarantee the real-time processing and low memory requirements by locating the area where targets may exist in advance. Different from former detection methods, multi-scan video data are utilized. In the stage of fine detection, the candidate areas are grouped into three categories: single target, dense targets and sea clutter. Different approaches for processing the different categories are implemented to achieve excellent performance. The superiority of the proposed framework beyond state-of-the-art baselines is well substantiated in this work. Low memory requirement of the proposed framework was verified by theoretical analysis. Real-time processing capability was verified by the video data of two real scenarios. Synthetic data were tested to show the improvement in tracking performance by using the proposed detection framework.

A Memory-Efficient Parallelizable Method for Computation of Thévenin Equivalents Used in Real-Time Stability Assessment

This paper introduces a factor-solve method that efficiently computes Thevenin equivalents for all buses in the power system. A range of real-time stability assessment methods relies on Thevenin equivalents, and it is therefore essential that these methods can be determined fast and efficiently. The factor-solve method has runtime for computing Thevenin voltage that scales linearly with system size resulting in runtime of only a few milliseconds even for systems with several thousand buses. The computations only need the sparse admittance matrix for the power system and a sparse factorization resulting in low memory requirements, and furthermore, Thevenin impedances can be determined in parallel. The factor-solve method is compared to a reference method that uses coefficients for super-position to determine the Thevenin equivalents. The reference method is shown to have unsatisfactory runtime and complexity. The factor-solve method is tested, parallelized, and analyzed, which shows a considerable speedup in computations of Thevenin equivalents enabling them to be computed in real time.

Zonal disturbance region update method for steady compressible viscous flows

The zonal disturbance region update method (zDRUM) presented in this work is an extension of the disturbance region update method (DRUM) to steady compressible viscous flows, capable of achieving convergence acceleration together with lower memory requirements. In the frame of DRUM and the zonal method, the new methodology taking advantage of the characteristics of the time-marching solution process employs two time-dependent dynamic computational domains where solely disturbed cells with non-convergent solutions are updated while the inviscid and the viscous flows are treated separately. A new data structure inspired by the pin art is introduced to store the dynamic computational domains more efficiently. Numerical results of six test cases in a wide range of Mach and Reynolds numbers demonstrate that, firstly, zDRUM accomplishes remarkable convergence speed for solving all compressible viscous flow problems, benefiting from the reduction in the computational effort per iteration; secondly, it is equally robust and efficient for different dimensions and flow types, for various reconstruction, spatial discretization and time-marching schemes; thirdly, it may reduce the maximum memory requirements.

Robust and parallel scalable iterative solutions for large-scale finite cell analyses

The finite cell method is a flexible discretization technique for numerical analysis on domains with complex geometries. By using a non-boundary conforming computational domain that can be easily meshed, automatized computations on a wide range of geometrical models can be performed. The application of the finite cell method, and other immersed methods, to large real-life and industrial problems is often limited due to the conditioning problems associated with these methods. These conditioning problems have caused researchers to resort to direct solution methods. This significantly limits the maximum size of solvable systems. Iterative solvers are better suited for large-scale computations than their direct counterparts due to their lower memory requirements and suitability for parallel computing. These benefits can, however, only be exploited when systems are properly conditioned. In this contribution we present an Additive-Schwarz type preconditioner that enables efficient and parallel scalable iterative solutions of large-scale multi-level hp-refined finite cell systems.

Low memory block tree coding for hyperspectral images

Hyperspectral image sensors are resource constrained and have limited on-board memory. Processing of high volume hyperspectral images pose a challenge to the memory and resources of the sensor. Contemporary wavelet based image compression schemes have intensive memory requirement of which 3D-WBTC have superior coding performance due to through the exploitation of the inter sub-band & intra sub-band redundancy. This paper presents a low memory implementation of 3D-WBTC which is a listless scheme by using the fixed size state memory to keep track of block set partitioning and significance testing of wavelet coefficients of transformed hyperspectral images. Memory access time is significantly reduced due to the elimination of lists that leads to reduced complexity. Simulation results of the proposed scheme shows that proposed coder is fast and have very low memory requirement thereby making it a suitable candidate for implementation in resource constrained hyper spectral image sensor.

Scattering of plane P- and SV-waves by periodic topography: Modeled by a PIBEM

For model of periodic topography subjected to plane waves, the wave fields have feature of repeating themselves with a certain shift of phase in the frequency domain. By fully exploring this particular feature, a periodic indirect boundary element method (PIBEM) is proposed to study the scattering of plane P- and SV-waves by periodic topography. The discretization effort of the PIBEM is reduced to a single topography using Green's functions of equivalent distributed loads acting on an inclined line as fundamental solutions. Compared to the traditional way of choosing a certain number of topographies to solve the problem with boundary conditions of other topographies being relaxed, the new method has the merits of higher precision and lower memory requirement. By taking periodic hill and canyon as examples, parametric studies are conducted to investigate the complex effects due to the periodic irregularity. Numerical results show that responses of periodic topography are quite different from those of a single and multiple topographies, demonstrating that it is very difficult to obtain the accurate results by choosing a certain number of topographies. In addition, periodic canyon has a stronger shielding effect on P-waves, while periodic hill has a stronger shielding effect on SV-waves.

A Fast Global Interpolation Method for Digital Terrain Model Generation from Large LiDAR-Derived Data

Airborne light detection and ranging (LiDAR) datasets with a large volume pose a great challenge to the traditional interpolation methods for the production of digital terrain models (DTMs). Thus, a fast, global interpolation method based on thin plate spline (TPS) is proposed in this paper. In the methodology, a weighted version of finite difference TPS is first developed to deal with the problem of missing data in the grid-based surface construction. Then, the interpolation matrix of the weighted TPS is deduced and found to be largely sparse. Furthermore, the values and positions of each nonzero element in the matrix are analytically determined. Finally, to make full use of the sparseness of the interpolation matrix, the linear system is solved with an iterative manner. These make the new method not only fast, but also require less random-access memory. Tests on six simulated datasets indicate that compared to recently developed discrete cosine transformation (DCT)-based TPS, the proposed method has a higher speed and accuracy, lower memory requirement, and less sensitivity to the smoothing parameter. Real-world examples on 10 public and 1 private dataset demonstrate that compared to the DCT-based TPS and the locally weighted interpolation methods, such as linear, natural neighbor (NN), inverse distance weighting (IDW), and ordinary kriging (OK), the proposed method produces visually good surfaces, which overcome the problems of peak-cutting, coarseness, and discontinuity of the aforementioned interpolators. More importantly, the proposed method has a similar performance to the simple interpolation methods (e.g., IDW and NN) with respect to computing time and memory cost, and significantly outperforms OK. Overall, the proposed method with low memory requirement and computing cost offers great potential for the derivation of DTMs from large-scale LiDAR datasets.

Efficient geometric integrators for nonadiabatic quantum dynamics. II. The diabatic representation.

Exact nonadiabatic quantum evolution preserves many geometric properties of the molecular Hilbert space. In the first paper of this series ["Paper I," S. Choi and J. Vaníček, J. Chem. Phys. 150, 204112 (2019)], we presented numerical integrators of arbitrary-order of accuracy that preserve these geometric properties exactly even in the adiabatic representation, in which the molecular Hamiltonian is not separable into kinetic and potential terms. Here, we focus on the separable Hamiltonian in diabatic representation, where the split-operator algorithm provides a popular alternative because it is explicit and easy to implement, while preserving most geometric invariants. Whereas the standard version has only second-order accuracy, we implemented, in an automated fashion, its recursive symmetric compositions, using the same schemes as in Paper I, and obtained integrators of arbitrary even order that still preserve the geometric properties exactly. Because the automatically generated splitting coefficients are redundant, we reduce the computational cost by pruning these coefficients and lower memory requirements by identifying unique coefficients. The order of convergence and preservation of geometric properties are justified analytically and confirmed numerically on a one-dimensional two-surface model of NaI and a three-dimensional three-surface model of pyrazine. As for efficiency, we find that to reach a convergence error of 10-10, a 600-fold speedup in the case of NaI and a 900-fold speedup in the case of pyrazine are obtained with the higher-order compositions instead of the second-order split-operator algorithm. The pyrazine results suggest that the efficiency gain survives in higher dimensions.