Reduce Memory Usage Research Articles

Multi-GPUs DEM algorithm and its application in the simulation of granular materials

The Discrete Element Method (DEM) is one of the most common numerical approach for modeling the physical and mechanical properties of granular materials. However, its computational efficiency can be a major bottleneck, which significantly hinders its application in engineering fields. To overcome this challenge, a multi-GPU framework with OpenMP has been developed. In this framework, the DEM particles in 3D space and their corresponding contact information are sorted in a given direction in the simulation, and each GPU device handles a subdomain of particles and relative contacts. Additionally, the contact information is densely arranged to minimize memory usage. To further enhance efficiency, sorting and packing algorithms for the contact information have been introduced. During calculating loop, the number of contacts on each device is monitored, and the spatial division of devices is adjusted to balance memory distribution. The efficacy of the multi-GPU algorithm is verified through a chute flow test. Using numerical results, the segregation process and mechanics of granular materials during sliding have been studied in detail. Furthermore, the simulations demonstrate that the developed multi-GPU algorithm greatly enhances computational efficiency and reduces memory usage per device.

Optimization of VR Application Mahakarya Vokasi Land Using Level of Detail (LOD) Technique

The development of the VR exhibition application for mahakarya vokasi Land using Level of Detail (LOD) is the focus of this research. LOD allows for adjusting the detail of 3D objects based on the user's viewing distance, reducing computational load, and improving application performance. LOD Groups are used to categorize objects based on complexity levels and dynamically adjust the level of detail. The research results indicate that the use of LOD successfully increased FPS by 284% and reduced memory usage in several important parameters such as setPass Calls, draw calls, batches, triangles, vertices, used textures, render textures, and shadow casters. Although some parameters did not show a significant improvement in FPS, it is important to monitor and optimize them. The use of LOD can be an effective strategy to improve the performance of VR exhibition applications by reducing rendering and memory load, creating a smoother and more responsive VR experience

Advancing Robot Perception in Non-Spiral Environments through Camera-based Image Processing

Robot perception heavily relies on camera-based visual input for navigating and interacting in its environment. As robots become integral parts of various applications, the need to efficiently compute their visibility regions in complex environments has grown. The key challenge addressed in this paper is to devise an innovative solution that not only accurately computes the visibility region V of a robot operating in a polynomial environment but also optimizes memory utilization to ensure real-time performance and scalability.The main objective of this research is to propose an algorithm that achieves optimal-time complexity and significantly reduces memory requirements for visibility region computation. By focusing on sub-linear memory utilization, we aim to enhance the robot's ability to perceive its surroundings effectively and efficiently.Previous approaches have provided solutions for visibility region computation in non-spiral environments, but most were not tailored to memory limitations. In contrast, the proposed algorithm is designed to achieve optimal time complexity that is O(n) while reducing memory usage to O(c/log n) variables, where c < n represents the number of critical corners in the environment. Leveraging the constant-memory model and memory-constrained algorithm, we aim to strike a balance between computational efficiency and memory usage.The algorithm's performance is rigorously evaluated through extensive simulations and practical experiments. The results demonstrate its linear-time complexity and substantial reduction in memory usage without compromising the accuracy of the visibility region computation. By efficiently handling memory constraints, the robot gains a cost-effective and reliable perception mechanism, making it well-suited for a wide range of real-world applications.The constant-memory model and memory-constrained algorithm presented in this paper offer a significant advancement in robot perception capabilities. By optimizing the visibility region computation in polynomial environments, our approach contributes to the efficient operation of robots, enhancing their performance and applicability in complex real-world scenarios. The results of this research hold promising potential for future developments in robotics, computer vision, and related fields.

Even Faster Exact Exchange for Solids via Tensor Hypercontraction.

Hybrid density functional theory (DFT) remains intractable for large periodic systems due to the demanding computational cost of exact exchange. We apply the tensor hypercontraction (THC) (or interpolative separable density fitting) approximation to periodic hybrid DFT calculations with Gaussian-type orbitals using the Gaussian plane wave approach. This is done to lower the computational scaling with respect to the number of basis functions (N) and k-points (Nk) at a fixed system size. Additionally, we propose an algorithm to fit only occupied orbital products via THC (i.e., a set of points, NISDF) to further reduce computation time and memory usage. This algorithm has linear scaling cost with k-points, no explicit dependence of NISDF on basis set size, and overall cubic scaling with unit cell size. Significant speedups and reduced memory usage may be obtained for moderately sized k-point meshes, with additional gains for large k-point meshes. Adequate accuracy can be obtained using THC-oo-K for self-consistent calculations. We perform illustrative hybrid density function theory calculations on the benzene crystal in the basis set and thermodynamic limits to highlight the utility of this algorithm.

A Smart Control System for the Oil Industry Using Text-to-Speech Synthesis Based on IIoT

Oil refineries have high operating expenses and are often exposed to increased asset integrity risks and functional failure. Real-time monitoring of their operations has always been critical to ensuring safety and efficiency. We proposed a novel Industrial Internet of Things (IIoT) design that employs a text-to-speech synthesizer (TTS) based on neural networks to build an intelligent extension control system. We enhanced a TTS model to achieve high inference speed by employing HiFi-GAN V3 vocoder in the acoustic model FastSpeech 2. We experimented with our system on a low resources-embedded system in a real-time environment. Moreover, we customized the TTS model to generate two target speakers (female and male) using a small dataset. We performed an ablation analysis by conducting experiments to evaluate the performance of our design (IoT connectivity, memory usage, inference speed, and output speech quality). The results demonstrated that our system Real-Time Factor (RTF) is 6.4 (without deploying the cache mechanism, which is a technique to call the previously synthesized speech sentences in our system memory). Using the cache mechanism, our proposed model successfully runs on a low-resource computational device with real-time speed (RTF equals 0.16, 0.19, and 0.29 when the memory has 250, 500, and 1000 WAV files, respectively). Additionally, applying the cache mechanism has reduced memory usage percentage from 16.3% (for synthesizing a sentence of ten seconds) to 6.3%. Furthermore, according to the objective speech quality evaluation, our TTS model is superior to the baseline TTS model.

Design and Implementation of Energy-Efficient Binary Neural Networks Using Adiabatic Quantum-Flux-Parametron Logic

Adiabatic quantum-flux-parametron (AQFP) logic is a promising technology for future energy-efficient, highperformance information processing systems. It has significantly low power dissipation thanks to the adiabatic switching of Josephson junctions. In this paper, we introduce an AQFPbased binary neural network (BNN) design methodology utilizing an in-memory computing scheme, analog accumulation, and a crossbar structure. The proposed design can effectively resolve the memory issue in superconducting digital circuits by significantly reducing memory usage in a non-Von Neumann fashion compared to conventional neural networks. As a proof-ofconcept, we designed and implemented an 8×8 AQFP BNN using the proposed design methodology targeted the AIST 10kA/cm2 4-layer niobium process. The Josephson junction count and energy dissipation of the proposed 8×8 BNN design are 2236 and 11.18 aJ, respectively.

High-Performance and Lightweight AI Model for Robot Vacuum Cleaners with Low Bitwidth Strong Non-Uniform Quantization

Artificial intelligence (AI) plays a critical role in the operation of robot vacuum cleaners, enabling them to intelligently navigate to clean and avoid indoor obstacles. Due to limited computational resources, manufacturers must balance performance and cost. This necessitates the development of lightweight AI models that can achieve high performance. Traditional uniform weight quantization assigns the same number of levels to all weights, regardless of their distribution or importance. Consequently, this lack of adaptability may lead to sub-optimal quantization results, as the quantization levels do not align with the statistical properties of the weights. To address this challenge, in this work, we propose a new technique called low bitwidth strong non-uniform quantization, which largely reduces the memory footprint of AI models while maintaining high accuracy. Our proposed non-uniform quantization method, as opposed to traditional uniform quantization, aims to align with the actual weight distribution of well-trained neural network models. The proposed quantization scheme builds upon the observation of weight distribution characteristics in AI models and aims to leverage this knowledge to enhance the efficiency of neural network implementations. Additionally, we adjust the input image size to reduce the computational and memory demands of AI models. The goal is to identify an appropriate image size and its corresponding AI models that can be used in resource-constrained robot vacuum cleaners while still achieving acceptable accuracy on the object classification task. Experimental results indicate that when compared to the state-of-the-art AI models in the literature, the proposed AI model achieves a 2-fold decrease in memory usage from 15.51 MB down to 7.68 MB while maintaining the same accuracy of around 93%. In addition, the proposed non-uniform quantization model reduces memory usage by 20 times (from 15.51 MB down to 0.78 MB) with a slight accuracy drop of 3.11% (the classification accuracy is still above 90%). Thus, our proposed high-performance and lightweight AI model strikes an excellent balance between model complexity, classification accuracy, and computational resources for robot vacuum cleaners.

Layered mixed-precision training: A new training method for large-scale AI models

How to efficiently and quickly train large-scale AI models has become a hot topic in recent deep learning. Mixed-precision training is an effective technique to speed up training and reduce memory usage. At present, the automatic mixed-precision training method mainly uses half-precision (FP16) for the matrix operations of forward and backward propagation of the entire model and accumulates the FP32 weight copies to avoid rounding errors. However, this method is not optimized for each layer individually, leading to poor convergence in large-scale model training because different layers have different data patterns. Therefore, this paper proposes a layered mixed-precision training method, which can flexibly adjust training precisions according to the contribution of each layer to the training effect. Applying the layered mixed-precision method, the ResNet model achieves a 1.9× speedup compared to the baseline and a lower percentage of accuracy loss. In addition, this paper combines the layered mixed-precision method with distributed training strategies. Combining data parallel training, the model achieves a 3.74× speedup by using four Tesla V100 GPUs. The applicability of the layered mixed-precision method in model parallel training has been verified. Combining optimized pipeline parallel training, the model achieves a 3.26× speedup by using three Tesla V100 GPUs.

One-Shot Replay: Boosting Incremental Object Detection via Retrospecting One Object

Modern object detectors are ill-equipped to incrementally learn new emerging object classes over time due to the well-known phenomenon of catastrophic forgetting. Due to data privacy or limited storage, few or no images of the old data can be stored for replay. In this paper, we design a novel One-Shot Replay (OSR) method for incremental object detection, which is an augmentation-based method. Rather than storing original images, only one object-level sample for each old class is stored to reduce memory usage significantly, and we find that copy-paste is a harmonious way to replay for incremental object detection. In the incremental learning procedure, diverse augmented samples with co-occurrence of old and new objects to existing training data are generated. To introduce more variants for objects of old classes, we propose two augmentation modules. The object augmentation module aims to enhance the ability of the detector to perceive potential unknown objects. The feature augmentation module explores the relations between old and new classes and augments the feature space via analogy. Extensive experimental results on VOC2007 and COCO demonstrate that OSR can outperform the state-of-the-art incremental object detection methods without using extra wild data.

MobileTL: On-Device Transfer Learning with Inverted Residual Blocks

Transfer learning on edge is challenging due to on-device limited resources. Existing work addresses this issue by training a subset of parameters or adding model patches. Developed with inference in mind, Inverted Residual Blocks (IRBs) split a convolutional layer into depthwise and pointwise convolutions, leading to more stacking layers, e.g., convolution, normalization, and activation layers. Though they are efficient for inference, IRBs require that additional activation maps are stored in memory for training weights for convolution layers and scales for normalization layers. As a result, their high memory cost prohibits training IRBs on resource-limited edge devices, and making them unsuitable in the context of transfer learning. To address this issue, we present MobileTL, a memory and computationally efficient on-device transfer learning method for models built with IRBs. MobileTL trains the shifts for internal normalization layers to avoid storing activation maps for the backward pass. Also, MobileTL approximates the backward computation of the activation layer (e.g., Hard-Swish and ReLU6) as a signed function which enables storing a binary mask instead of activation maps for the backward pass. MobileTL fine-tunes a few top blocks (close to output) rather than propagating the gradient through the whole network to reduce the computation cost. Our method reduces memory usage by 46% and 53% for MobileNetV2 and V3 IRBs, respectively. For MobileNetV3, we observe a 36% reduction in floating-point operations (FLOPs) when fine-tuning 5 blocks, while only incurring a 0.6% accuracy reduction on CIFAR10. Extensive experiments on multiple datasets demonstrate that our method is Pareto-optimal (best accuracy under given hardware constraints) compared to prior work in transfer learning for edge devices.

Detecting unregistered users through semi-supervised anomaly detection with similarity datasets

Recent research has focused on exploring systems that incorporate anomaly detection models to automate the addition of users in user recognition systems. Anomaly detection, a method used to distinguish between new and existing users by identifying abnormal images, has gained significant attention. Researchers have been actively investigating the Semi-Supervised Learning method, which utilizes only existing user data to differentiate between existing and new users. However, existing semi-supervised learning based anomaly detection models exhibit high performance on datasets with low similarity but experience a sharp decline in performance on datasets with high similarity. Furthermore, their large model size makes it challenging to execute them on edge nodes. To address these limitations, this paper proposes a model that can be executed on edge nodes and guarantees good performance on both low and high similarity datasets. The proposed model utilizes the LeNet-5, a user recognition model with fewer weights and multiple images as input, for classifying new users. This study compared the existing anomaly detection models with the proposed model using three datasets with varying similarities. The performance evaluation involved comparing the accuracy, ROC curve, and AUC of each model on a training server. Subsequently, the top three models were optimized for execution on the edge node (STM32F207ZG MCU) and further evaluated by comparing their accuracy, inference speed, and model size. The results revealed that the proposed model achieved an approximate 53% improvement in accuracy compared to the existing anomaly detection models. Furthermore, when executed on the edge node, the proposed model demonstrated significant memory savings, with a maximum reduction of approximately 530% and approximately 40% reduction in flash memory usage compared to the existing models.

Computational Integral Imaging Reconstruction via Elemental Image Blending without Normalization

This paper presents a novel computational integral imaging reconstruction (CIIR) method using elemental image blending to eliminate the normalization process in CIIR. Normalization is commonly used in CIIR to address uneven overlapping artifacts. By incorporating elemental image blending, we remove the normalization step in CIIR, leading to decreased memory consumption and computational time compared to those of existing techniques. We conducted a theoretical analysis of the impact of elemental image blending on a CIIR method using windowing techniques, and the results showed that the proposed method is superior to the standard CIIR method in terms of image quality. We also performed computer simulations and optical experiments to evaluate the proposed method. The experimental results showed that the proposed method enhances the image quality over that of the standard CIIR method, while also reducing memory usage and processing time.

Recursive State Machine Guided Graph Folding for Context-Free Language Reachability

Context-free language reachability (CFL-reachability) is a fundamental framework for program analysis. A large variety of static analyses can be formulated as CFL-reachability problems, which determines whether specific source-sink pairs in an edge-labeled graph are connected by a reachable path, i.e., a path whose edge labels form a string accepted by the given CFL. Computing CFL-reachability is expensive. The fastest algorithm exhibits a slightly subcubic time complexity with respect to the input graph size. Improving the scalability of CFL-reachability is of practical interest, but reducing the time complexity is inherently difficult. In this paper, we focus on improving the scalability of CFL-reachability from a more practical perspective---reducing the input graph size. Our idea arises from the existence of trivial edges, i.e., edges that do not affect any reachable path in CFL-reachability. We observe that two nodes joined by trivial edges can be folded---by merging the two nodes with all the edges joining them removed---without affecting the CFL-reachability result. By studying the characteristic of the recursive state machines (RSMs), an alternative form of CFLs, we propose an approach to identify foldable node pairs without the need to verify the underlying reachable paths (which is equivalent to solving the CFL-reachability problem). In particular, given a CFL-reachability problem instance with an input graph G and an RSM, based on the correspondence between paths in G and state transitions in RSM, we propose a graph folding principle, which can determine whether two adjacent nodes are foldable by examining only their incoming and outgoing edges. On top of the graph folding principle, we propose an efficient graph folding algorithm GF. The time complexity of GF is linear with respect to the number of nodes in the input graph. Our evaluations on two clients (alias analysis and value-flow analysis) show that GF significantly accelerates RSM/CFL-reachability by reducing the input graph size. On average, for value-flow analysis, GF reduces 60.96% of nodes and 42.67% of edges of the input graphs, obtaining a speedup of 4.65× and a memory usage reduction of 57.35%. For alias analysis, GF reduces 38.93% of nodes and 35.61% of edges of the input graphs, obtaining a speedup of 3.21× and a memory usage reduction of 65.19%.

Quantization aware training for efficient reinforcement learning network

In recent years, reinforcement learning has been applied to many areas, especially deep reinforcement learning. Deep learning-based methods use a large computing method known as Quantization Aware Training (QAT). For this project, QAT will be used to investigate the performance of Deep Q-Networks. This method of QAT can reduce memory usage of the network, but may affect the training performance of the network with some games.

Superpixel-based linear reconstruction method for dual-camera compressed hyperspectral imaging system

Compressive spectral imaging systems have received significant research attention in recent years. Among these systems, the dual-camera compressive hyperspectral imaging (DCCHI) design, has the potential to improve imaging resolution while maintaining the snapshot advantage. In this paper, we propose superpixel-based linear reconstruction methods for DCCHI system. Specifically, we assume that the spectral intensity of the superpixels generated from the RGB image varies linearly in the spectral dimension. Instead of addressing an ill-posed problem, we reconstruct hyperspectral images via solving overdetermined equations using the least squares method, resulting in a significant reduction in both computation time and memory usage. To further enhance the reconstruction accuracy, we extend the spectra of superpixels from a 1D linear space into a 3D space with an RGB spectral response matrix. Despite the simplicity of our method, it offers improved reconstruction accuracy of hyperspectral images compared to other state-of-the-art methods. Furthermore, it is more efficient in terms of memory usage and computational time.

Hardware architecture design for real-time SIFT extraction with reduced memory usage

Hardware architecture design for real-time SIFT extraction with reduced memory usage

C-Coupler3.0: an integrated coupler infrastructure for Earth system modelling

Abstract. The community coupler (C-Coupler) for Earth system modelling is a coupler family that was developed in China in 2010. C-Coupler3.0, the latest version, is fully compatible with the previous version, C-Coupler2, and is an integrated infrastructure with new features, i.e. a series of parallel-optimization technologies for accelerating coupling initialization and reducing memory usage, a common halo-exchange library for developing a parallel version of a model, a common module-integration framework for integrating a software module (e.g. a flux algorithm, a parameterization scheme, and a data assimilation method), a common framework for conveniently developing a weakly coupled ensemble data assimilation system, and a common framework for flexibly inputting and outputting fields in parallel. Specifically, C-Coupler3.0 is able to handle coupling under much finer resolutions (e.g. more than 100 million horizontal grid cells) with fast coupling initialization and successful generation of remapping-weight files.

LazySort: A customized sorting algorithm for non-volatile memory

Due to the development of new non-volatile storage technologies, bridging the access delay gap between storage and memory has been promoted. Therefore, since the volume of data continues to increase, the need for new types of non-volatile storage for high-performance computing and database management systems significantly increases. As a basic algorithm of the database management system, external sorting is widely used but suffers from numerous I/O operations caused by the frequent data exchange between the memory and external memory. Non-volatile memory (NVM) is a promising solution. However, the traditional external sorting algorithm is only designed for disk storage and cannot be used for NVM. To solve this issue, we propose LazySort, which is a novel external sorting algorithm that exploits the NVM byte addressing mechanism and the orderly distribution of data to improve the sorting efficiency. Specifically, LazySort first detects the ordered data segment in the memory and then records the storage location of the data segment as well as its maximum/minimum values in an index table named ITable. Next, to reduce memory usage and accelerate the sorting procedure, LazySort performs an optimization strategy RunMerge to merge non-intersecting data blocks according to the range of ITable records. To verify the performance of LazySort, we built an experimental platform with the NVM-DRAM storage architecture combining NVM and dynamic random access memory (DRAM) and conducted a series of experiments. The experimental results show that LazySort is more efficient than the traditional external sorting algorithm, since the sorting time of LazySort is shortened by 93.08% and the number of NVM write operations is reduced by 49.50%. Moreover, the advantage of LazySort over traditional algorithms is more significant when the data amount increases.

Angular-spatial discontinuous Galerkin method for radiative heat transfer with a participating medium in complex three-dimensional geometries

The angular-spatial discontinuous Galerkin method (ASDGM) is a numerical technique developed to solve the radiative transfer equation (RTE) in complex three-dimensional geometries with absorbing, emitting, and scattering media. The ASDGM discretizes both the angular and spatial domains of the RTE using structured quadrilateral elements for the angular domain and unstructured tetrahedral elements for the spatial domain. In the solution process, the Riemann upwind solver separates the input and output information at the tetrahedral element boundary. Moreover, the discontinuous Galerkin (DG) sweep algorithm optimizes the order of the unstructured tetrahedral elements and reduces memory usage caused by the coupling of the discretization in the angular and spatial domains. The performance of the ASDGM has been evaluated using several radiative transfer cases in complex three-dimensional geometries, which demonstrate its good accuracy compared to available data from the literature. Additionally, the computational time in the angular domain for high-order DG is lower than that of the finite element method for the same discrete angles. Furthermore, the DG sweep algorithm provides a faster convergence rate in the spatial domain compared to the original sweep algorithm. Therefore, the ASDGM with the DG sweep algorithm is a preferable method for solving radiative transfer problems in complex three-dimensional geometries.

Class-Incremental Learning: Survey and Performance Evaluation on Image Classification.

For future learning systems, incremental learning is desirable because it allows for: efficient resource usage by eliminating the need to retrain from scratch at the arrival of new data; reduced memory usage by preventing or limiting the amount of data required to be stored - also important when privacy limitations are imposed; and learning that more closely resembles human learning. The main challenge for incremental learning is catastrophic forgetting, which refers to the precipitous drop in performance on previously learned tasks after learning a new one. Incremental learning of deep neural networks has seen explosive growth in recent years. Initial work focused on task-incremental learning, where a task-ID is provided at inference time. Recently, we have seen a shift towards class-incremental learning where the learner must discriminate at inference time between all classes seen in previous tasks without recourse to a task-ID. In this paper, we provide a complete survey of existing class-incremental learning methods for image classification, and in particular, we perform an extensive experimental evaluation on thirteen class-incremental methods. We consider several new experimental scenarios, including a comparison of class-incremental methods on multiple large-scale image classification datasets, an investigation into small and large domain shifts, and a comparison of various network architectures.