Reduce Memory Requirements Research Articles

HETEROGENEOUS COMPUTING TO ACCELERATE THE SEARCH OF SUPER K-MERS BASED ON MINIMIZERS

The k-mers processing techniques based on partitioning of the data set on the disk using minimizer-type seeds have led to a significant reduction in memory requirements; however, it has added processes (search and distribution of super k-mers) that can be intensive given the large volume of data. This paper presents a massive parallel processing model in order to enable the efficient use of heterogeneous computation to accelerate the search of super k-mers based on seeds (minimizers or signatures). The model includes three main contributions: a new data structure called CISK for representing the super k-mers, their minimizers and two massive parallelization patterns in an indexed and compact way: one for obtaining the canonical m-mers of a set of reads and another for searching for super k-mers based on minimizers. The model was implemented through two OpenCL kernels. The evaluation of the kernels shows favorable results in terms of execution times and memory requirements to use the model for constructing heterogeneous solutions with simultaneous execution (workload distribution), which perform co-processing using the current search methods of super k -mers on the CPU and the methods presented herein on GPU. The model implementation code is available in the repository: https://github.com/BioinfUD/K-mersCL.

Modified REP Pattern for 3×3 Kernel Polar Codes

Using large size kernels in conjunction with the binary kernel enables length flexibility in polar codes. However, the resultant multi-kernel polar codes suffer from high decoding complexity. The size-3 kernels have drawn much attention due to their sufficiently high polarization exponents and the lowest decoding complexity among larger than size-2 kernels. However, the two specific size-3 kernels that can achieve the optimal polarization have a zero in their last rows. This results in the error-rate performance degradation of the repetition (REP) and generalized-repetition (G-REP) nodes in addition to increased memory requirements for their fast decoders. In this letter, we propose modifications to the REP and G-REP nodes that simultaneously result in improved error-rate performance and reduced memory requirements for their fast decoders.

An enhanced color visual cryptography scheme based on the binary dragonfly algorithm

ABSTRACT Visual cryptography is an encryption technique that decomposes a secret image into multiple shares that are digitally ‘overlapped’ based on logical operations to recover the original secret image. Each individual share cannot reveal clues about the secret image. Existing visual cryptography schemes suffer from problems such as pixel expansion, high computational complexity and poor decryption quality. To overcome these problems, we propose a new 2-out-of-2 secret sharing scheme for color images based on the binary dragonfly algorithm. The dragonfly algorithm is used to determine optimal color levels for the encryption process, which leads to reconstructed images with the higher quality upon decryption with minimal effect on the overall computational complexity. The proposed scheme is also non-expansible, whereby the share images are of the same size as the original secret image. This leads to reduced memory requirements as well as improved image quality. We analyze the strength of the proposed scheme based on a standard set of metrics which includes an evaluation of its resistance against cryptanalytic attacks, histogram, correlation, entropy and encryption quality. Results show that the proposed visual cryptography scheme has desired properties such as improved recovered image quality, near-ideal statistical characteristics and fast encryption speed, depicting an improvement over existing schemes.

An Efficient Parallel Hybrid Method of FEM-MLFMA for Electromagnetic Radiation and Scattering Analysis of Separated Objects

In order to meet the rapidly increasing demand for accurate and efficient analysis of complex radiating or scattering structures in the presence of electrically large objects, a finite element method (FEM)-multilevel fast multipole algorithm (MLFMA) hybrid method that based on the Finite Element-Iterative Integral Equation Evaluation (FE-IIEE) mesh truncation technique is proposed in this paper. The present method makes use of FEM for the regions with small and complex features and MLFMA for the analysis of the electrically large objects, which ensure the accuracy and applicability of the method are better than most commonly adopted FEM-high frequency technique (HFT) hybrid method. The mutual interactions between regions are taken into account in a fully coupled way through iterative near filed computation process. In order to achieve an excellent performance, both algorithms have been implemented together from scratch, being able to run over multi CPU cores. An efficient parallel FEM domain decomposition method (DDM) solver with exploiting geometrical repetitions is included to drastically reduce memory requirements and computational time in the calculation of large array antenna. Also, the parallel MLFMA is adopted to expedite the near-field information exchange between regions. Through numerical example, the effect of distance between regions on the convergence of the proposed hybrid method is studied, and it is shown that the proposed method converge well even if the distance is equal to 0.05λ. Through comparisons with an in-house higher order method of moments (HOMoM) code and commercial software FEKO, the accuracy and effectiveness of the implemented parallel hybrid method are validated showing excellent performance.

Photo-acoustic tomographic image reconstruction from reduced data using physically inspired regularization

We propose a model-based image reconstruction method for photoacoustic tomography (PAT) involving a novel form of regularization and demonstrate its ability to recover good quality images from significantly reduced size datasets. The regularization is constructed to suit the physical structure of typical PAT images. We construct it by combining second-order derivatives and intensity into a non-convex form to exploit a structural property of PAT images that we observe: in PAT images, high intensities and high second-order derivatives are jointly sparse. The specific form of regularization constructed here is a modification of the form proposed for fluorescence image restoration. This regularization is combined with a data fidelity cost, and the required image is obtained as the minimizer of this cost. As this regularization is non-convex, the efficiency of the minimization method is crucial in obtaining artifact-free reconstructions. We develop a custom minimization method for efficiently handling this non-convex minimization problem. Further, as non-convex minimization requires a large number of iterations and the PAT forward model in the data-fidelity term has to be applied in the iterations, we propose a computational structure for efficient implementation of the forward model with reduced memory requirements. We evaluate the proposed method on both simulated and real measured data sets and compare them with a recent reconstruction method that is based on a well-known fast iterative shrinkage threshold algorithm (FISTA).

A Resource-Efficient Parallel Connected Component Labeling Algorithm and Its Hardware Implementation

Connected Component labeling (CCL) is usually time-consuming, so a dedicated hardware accelerator of CCL is essential in the embedded vision and multimedia system. In this paper, we propose a single-scan resource-efficient parallel CCL algorithm. Our CCL method scans two adjacent rows simultaneously to extract runs and detect equivalent runs; an equivalent label set is used to resolve equivalences. After each row scan, the finished objects are output in time, and the freed memory resources are reused to reduce memory requirements. Both pixel-based labeled image (PLI) and run-based labeled image (RLI) can be generated by our CCL method. In addition, the steps of our CCL method are executed concurrently to improve labeling performance. The hardware architecture based on our CCL method is implemented with Verilog. The evaluation results illustrate that our CCL architecture can label more than 40 2048 × 1536 benchmark images per second on average, and outperforms previous CCL architectures in terms of labeling performance or memory resource consumption.

Entropy-Driven Adaptive Filtering for High-Accuracy and Resource-Efficient FPGA-Based Neural Network Systems

Binarized neural networks are well suited for FPGA accelerators since their fine-grained architecture allows the creation of custom operators to support low-precision arithmetic operations, and the reduction in memory requirements means that all the network parameters can be stored in internal memory. Although good progress has been made to improve the accuracy of binarized networks, it can be significantly lower than networks where weights and activations have multi-bit precision. In this paper, we address this issue by adaptively choosing the number of frames used during inference, exploiting the high frame rates that binarized neural networks can achieve. We present a novel entropy-based adaptive filtering technique that improves accuracy by varying the system’s processing rate based on the entropy present in the neural network output. We focus on using real data captured with a standard camera rather than using standard datasets that do not realistically represent the artifacts in video stream content. The overall design has been prototyped on the Avnet Zedboard, which achieved 70.4% accuracy with a full processing pipeline from video capture to final classification output, which is 1.9 times better compared to the base static frame rate system. The main feature of the system is that while the classification rate averages a constant 30 fps, the real processing rate is dynamic and varies between 30 and 142 fps, adapting to the complexity of the data. The dynamic processing rate results in better efficiency that simply working at full frame rate while delivering high accuracy.

Modular Neural Networks for Low-Power Image Classification on Embedded Devices

Embedded devices are generally small, battery-powered computers with limited hardware resources. It is difficult to run deep neural networks (DNNs) on these devices, because DNNs perform millions of operations and consume significant amounts of energy. Prior research has shown that a considerable number of a DNN’s memory accesses and computation are redundant when performing tasks like image classification. To reduce this redundancy and thereby reduce the energy consumption of DNNs, we introduce the Modular Neural Network Tree architecture. Instead of using one large DNN for the classifier, this architecture uses multiple smaller DNNs (called modules ) to progressively classify images into groups of categories based on a novel visual similarity metric. Once a group of categories is selected by a module, another module then continues to distinguish among the similar categories within the selected group. This process is repeated over multiple modules until we are left with a single category. The computation needed to distinguish dissimilar groups is avoided, thus reducing redundant operations, memory accesses, and energy. Experimental results using several image datasets reveal the effectiveness of our proposed solution to reduce memory requirements by 50% to 99%, inference time by 55% to 95%, energy consumption by 52% to 94%, and the number of operations by 15% to 99% when compared with existing DNN architectures, running on two different embedded systems: Raspberry Pi 3 and Raspberry Pi Zero.

A Fast Parallel Solution Technique for Large Periodic Structures Based on FEM-DDM

In this letter, a fast solution technique based on the finite-element approximation and the domain decomposition method for the analysis of superlarge electromagnetic periodic structures is presented. This technique reduces memory requirements and computational time by decomposing the original periodic structure into several nonoverlapping subdomains according to geometrical repetitions and utilizing the characteristics that most of the subdomains will generate identical system matrices. An adaptive parallel framework considering periodic element features is designed to further accelerate the proposed technique. Through comparisons with existing methods and the commercial software ANSYS HFSS, the benefits and accuracy of the proposed technique are demonstrated. Finally, a microstrip patch antenna array with 10 000 elements and a large waveguide slot antenna array with 4096 slots are simulated showing the power of the presented method.

Fast image blending for high-quality panoramic images on mobile phones

This paper presents a fast image blending approach for combining a set of registered images into a composite mosaic with no visible seams and minimal texture distortion on mobile phones. A unique seam image is generated using two-pass nearest distance transform, which is independent on the order of input images and has good scalability. Each individual mask can be extracted from this seam image quickly. To promote blending speed and reduce memory usage in building high resolution image mosaics on mobile phones, the seam image and mask images are compressed using run-length encoding, and all the following mask operations are built on run-length encoding scheme. Moreover, single instruction multiple data instruction set is used in Gaussian and Laplacian pyramids construction to improve the blending speed further. The use of run-length encoding for masks processing leads to reduced memory requirements and a compact storage of the mask data, and the use of single instruction multiple data instruction set achieves better parallelism and faster execution speed on mobile phones.

A high-speed feature matching method of high-resolution aerial images

This paper presents a novel corner detection and scale estimation algorithm for image feature description and matching. Inspired by Adaboost’s weak classifier, a series of sub-detectors is elaborately designed to obtain reliable corner pixels. The new corner detection algorithm is more robust than the FAST and HARRIS algorithm, and it is especially suitable for the implementation in FPGA. The new scale estimation method can be directly implemented in the original image without building Gaussian pyramid and searching max response value in each level, which not only increase computational efficiency but also greatly reduces memory requirement. Based on the proposed algorithm, a CPU-FPGA cooperative parallel processing architecture is presented. The architecture overcomes the memory space limitation of FPGA and achieves high-speed feature matching for massive high-resolution aerial images. The speed of the CPU-FPGA cooperative process is hundred times faster than SIFT algorithm running on CPU, and dozens of times faster than SIFT running in CPU + GPU system.

Large eddy simulation of spray combustion using flamelet generated manifolds combined with artificial neural networks

In the present work, artificial neural networks (ANN) technique combined with flamelet generated manifolds (FGM) is proposed to mitigate the memory issue of FGM models. A set of ANN models is firstly trained using a 68-species mass fractions in mixture fraction-progress variable space. The ANN prediction accuracy is examined in large eddy simulation (LES) and Reynolds averaged Navier-Stokes (RANS) simulations of spray combustion. It is shown that the present ANN models can properly replicate the FGM table for most of the species mass fractions. The network models with relative error less than 5% are considered in RANS and LES to simulate the Engine Combustion Network (ECN) Spray H flames. Validation of the method is firstly conducted in the framework of RANS. Both non-reacting and reacting cases show the present method predicts very well the trend of spray and combustion process under different ambient temperatures. The results show that FGM-ANN can replicate the ignition delay time (IDT) and lift-off length (LOL) precisely as the conventional FGM method, and the results agree very well with the experiments. With the help of ANN, it is possible to achieve high efficiency and accuracy, with a significantly reduced memory requirement of the FGM models. LES with FGM-ANN is then applied to explore the detailed spray combustion process. Chemical explosive mode analysis (CEMA) approach is used to identify the local combustion modes. It is found that before the spray flame is developed to the steady-state, the high CH2O zone is always associated with ignition mode. However, high CH2O zone together with high OH zone is dominated by the burned mode after the steady-state. The lift-off position is dominated mainly by the diffusion mode.

A tutorial and open source software for the efficient evaluation of gravity and magnetic kernels

Fast computation of three-dimensional gravity and magnetic forward models is considered. When the measurement data is assumed to be obtained on a uniform grid which is staggered with respect to the discretization of the parameter volume, the resulting kernel sensitivity matrices exhibit block-Toeplitz–Toeplitz-block (BTTB) structure. These matrices are symmetric for the gravity problem but unsymmetric for the magnetic problem. In each case, the structure facilitates fast forward computation using two-dimensional fast Fourier transforms. The construction of the kernel matrices and the application of the transform for fast forward multiplication, for each problem, is carefully described. But, for purposes of comparison with the non-transform approach, the generation of the unique entries that define a given kernel matrix is also explained. It is also demonstrated how the matrices, and hence transforms, are adjusted when padding around the volume domain is introduced. The transform algorithms for fast forward matrix multiplication with the sensitivity matrix and its transpose, without the direct construction of the relevant matrices, are presented. Numerical experiments demonstrate the significant reduction in computation time and memory requirements that are achieved using the transform implementation. Thus, it becomes feasible, both in terms of reduced memory requirements and computational time, to implement the transform algorithms for large three-dimensional volumes. All presented algorithms, including with variable padding, are coded for optimal memory, storage and computation as an open source Matlab code which can be adapted for any convolution kernel which generates a BTTB matrix, whether or not it is symmetric. This work, therefore, provides a general tool for the efficient simulation of gravity and magnetic field data, as well as any formulation which admits a sensitivity matrix with the required structure.

Low Memory Architectures of Fractional Wavelet Filter for Low-Cost Visual Sensors and Wearable Devices

This paper proposes a low memory architecture of fractional wavelet filter for computing two-dimensional discrete wavelet transform of gray-scale images. The proposed architecture uses a data scanning order which mainly contributes to memory reduction. A multiplierless version of the proposed architecture is also designed to reduce the critical path delay equal to the delay of an adder. The proposed architectures and an existing low memory discrete wavelet transform architecture are implemented on the same field programmable gate array board and their performance are compared for different image resolutions. The results show that the proposed architectures (with and without multipliers) uses lesser basic blocks, such as look-up tables (about 85-94% less), flip-flops (about 85-90% less), consumes lesser dynamic power (about 58-75% less), and results in critical path delay reduction of (about 23-77%) compared to one of the best existing low memory discrete wavelet transform architecture for images of $1024\times {1024}$ pixels. These advantages are achieved in addition to reduction in memory requirement over the existing state-of-the-art architectures. Thus, the proposed architectures are viable options for calculating discrete wavelet transform coefficients of gray-scale images on low-cost memory-constrained platforms like visual sensors/wearable devices and Internet of things.

Neuromorphic Systems Design by Matching Inductive Biases to Hardware Constraints.

Neuromorphic systems are designed with careful consideration of the physical properties of the computational substrate they use. Neuromorphic engineers often exploit physical phenomena to directly implement a desired functionality, enabled by “the isomorphism between physical processes in different media” (Douglas et al., 1995). This bottom-up design methodology could be described as matching computational primitives to physical phenomena. In this paper, we propose a top-down counterpart to the bottom-up approach to neuromorphic design. Our top-down approach, termed “bias matching,” is to match the inductive biases required in a learning system to the hardware constraints of its implementation; a well-known example is enforcing translation equivariance in a neural network by tying weights (replacing vector-matrix multiplications with convolutions), which reduces memory requirements. We give numerous examples from the literature and explain how they can be understood from this perspective. Furthermore, we propose novel network designs based on this approach in the context of collaborative filtering. Our simulation results underline our central conclusions: additional hardware constraints can improve the predictions of a Machine Learning system, and understanding the inductive biases that underlie these performance gains can be useful in finding applications for a given constraint.

PGeoTopic: A Distributed Solution for Mining Geographical Topic Models

Geographical topic models have been used to mine geo-tagged documents for topical region and geographical topics, and also have applications in recommendations, user mobility modeling, event detection, etc. Existing studies focus on learning effective geographical topic models while ignoring the efficiency issue. However, it is very expensive to train geographical topic models — it may take days to train a geographical topic model of a small scale on a collection of documents with millions of word tokens. In this paper, we propose the first distributed solution, called <inline-formula><tex-math notation="LaTeX">${\sf PGeoTopic}$</tex-math></inline-formula> , for training geographical topic models. The proposed solution comprises several novel technical components to increase parallelism, reduce memory requirement, and reduce communication cost. Experiments show that our approach for mining geographical topic models is scalable with both model size and data size on distributed systems.

Weightless Neural Networks as Memory Segmented Bloom Filters

Weightless Neural Networks (WNNs) are Artificial Neural Networks based on RAM memory broadly explored as solution for pattern recognition applications. Memory-oriented solutions for pattern recognition are typically very simple, and can be easily implemented in hardware and software. Nonetheless, the straightforward implementation of a WNN requires a large amount of memory resources making its adoption impracticable on memory constrained systems. In this paper, we establish a foundational relationship between WNN and Bloom filters, presenting a novel unified framework which encompasses the two. In particular, we indicate that a WNN can be framed as a memory segmented Bloom filter. Leveraging such finding, we propose a new model of WNNs which utilizes Bloom filters to implement RAM nodes. Bloom filters reduce memory requirements, and allow false positives when determining if a given pattern was already seen in data. We experimentally found that for pattern recognition purposes such false positives can build robustness into the system. The experimental results show that our model using Bloom filters achieves competitive accuracy, training time and testing time, consuming up to 6 orders of magnitude less memory resources when compared against the standard Weightless Neural Network model.

Hardware-Compliant Compressive Image Sensor Architecture Based on Random Modulations and Permutations for Embedded Inference

This work presents a compact CMOS Image Sensor (CIS) architecture enabling embedded object recognition facilitated by a dedicated end-of-column Compressive Sensing (CS), reducing on-chip memory needs. Our sensing scheme is based on a combination of random modulations and permutations leading to an implementation with very limited hardware impacts. It is designed to meet both theoretical ( i.e. , stable embedding, measurements incoherence) and practical requirements ( i.e. , silicon footprint, power consumption). The only additional hardware compared to a standard CIS architecture using first order incremental Sigma-Delta ( $\Sigma \Delta $ ) Analog to Digital Converter (ADC) are a pseudo-random data mixing circuit, an in- $\Sigma \Delta ~\pm 1$ modulator and a small Digital Signal Processor (DSP). On the algorithmic side, three variants are presented to perform the inference on compressed measurements with a tunable complexity ( i.e. , one-vs.-all SVM, hierarchical SVM and small ANN with 1-D max-pooling). An object recognition accuracy of $\simeq 98.8$ % is reached on the COIL database (COIL, 100 classes) using our dedicated Neural Network classifier. We stress that the signal-independent dimensionality reduction performed by our dedicated CS scheme (1/480 in $480 \times 640$ VGA resolution case) allows to dramatically reduce memory requirements mainly related to the remotely learned coefficients used for the inference stage.

Acoustic Noise Classification Using Selective Discrete Wavelet Transform-Based Mel-Frequency Cepstral Coefficient

A feature extraction method through wavelet mel-frequency cepstral coefficients (MFCCs) is proposed for acoustic noise classification. The method combined with a wavelet sub-band selection technique and a feedforward neural network with two hidden layers, is a promising solution for a compact acoustic noise classification system that could be added to speech enhancement systems and deployed in hearing devices such as cochlear implants. The technique leads to higher classification accuracies (with a mean of 95.25%) across three SNR values, a significantly smaller feature set with 16 features, a reduced memory requirement, and faster training convergence, with a trade-off of slightly higher computational complexity by a factor of 1.89 in comparison to the traditional short-time Fourier transform-based (STFT-based) technique.

An efficient cell-centered finite-volume method with face-averaged nodal-gradients for triangular grids

In this paper, a face-averaged nodal-gradient approach is proposed as an efficient gradient method for a second-order cell-centered finite-volume discretization on triangular grids. The gradients needed in the linear reconstruction are computed in two steps: (1) compute gradients at nodes from solutions stored at cells, and (2) compute the gradient at each face by averaging the nodal gradients over the nodes forming the face. A unique feature of the proposed approach is that the same face-averaged gradient is used in linear reconstructions from both the left and right cells to the face centroid. This approach offers advantages over a conventional cell-gradient-based finite-volume method: reduced memory requirement for the gradients, smaller residual stencils, direct evaluations of viscous stresses and heat fluxes at solid boundaries, and less interpartition communication. The method is demonstrated for various inviscid and viscous problems from low to high Mach numbers on triangular grids.