Reduce Memory Usage Research Articles

MCMC: Multi-Constrained Model Compression via One-Stage Envelope Reinforcement Learning.

Model compression methods are being developed to bridge the gap between the massive scale of neural networks and the limited hardware resources on edge devices. Since most real-world applications deployed on resource-limited hardware platforms typically have multiple hardware constraints simultaneously, most existing model compression approaches that only consider optimizing one single hardware objective are ineffective. In this article, we propose an automated pruning method called multi-constrained model compression (MCMC) that allows for the optimization of multiple hardware targets, such as latency, floating point operations (FLOPs), and memory usage, while minimizing the impact on accuracy. Specifically, we propose an improved multi-objective reinforcement learning (MORL) algorithm, the one-stage envelope deep deterministic policy gradient (DDPG) algorithm, to determine the pruning strategy for neural networks. Our improved one-stage envelope DDPG algorithm reduces exploration time and offers greater flexibility in adjusting target priorities, enhancing its suitability for pruning tasks. For instance, on the visual geometry group (VGG)-16 network, our method achieved an 80% reduction in FLOPs, a 2.31× reduction in memory usage, and a 1.92× acceleration, with an accuracy improvement of 0.09% compared with the baseline. For larger datasets, such as ImageNet, we reduced FLOPs by 50% for MobileNet-V1, resulting in a 4.7× faster speed and 1.48× memory compression, while maintaining the same accuracy. When applied to edge devices, such as JETSON XAVIER NX, our method resulted in a 71% reduction in FLOPs for MobileNet-V1, leading to a 1.63× faster speed, 1.64× memory compression, and an accuracy improvement.

Accounting Management and Optimizing Production Based on Distributed Semantic Recognition

Accounting management and production optimization are vital aspects of enterprise management, serving as indispensable core components in the modern business landscape. However, conventional methods reliant on manual input exhibit drawbacks such as low recognition accuracy and excessive memory consumption. To address these challenges, semantic recognition technology utilizing voice signals has emerged as a pivotal solution across various industries. Building upon this premise, this paper introduces a distributed semantic recognition‐based algorithm for accounting management and production optimization. The proposed algorithm encompasses multiple modules, including a front‐end feature extraction module, a channel transmission module, and a voice quality vector quantization module. Additionally, a semantic recognition module is introduced to process the voice signals and generate prediction results. By leveraging extensive accounting management and production data for learning and analysis, the algorithm automatically uncovers patterns and laws within the data, extracting valuable information. To validate the proposed algorithm, this study utilizes the dataset from the UCI machine learning repository and applies it for analysis and processing. The experimental findings demonstrate that the algorithm introduced in this paper outperforms alternative methods. Specifically, it achieves a notable 9.3% improvement in comprehensive recognition accuracy and reduces memory usage by 34.4%. These results highlight the algorithm’s efficacy in enhancing the understanding and analysis of customer needs, market trends, competitors, and other pertinent information within the realm of commercial applications for companies.

Reducing the memory usage of structured grid solvers with wavelet compression

This paper presents a new solution to address the challenge of increasing memory usage in high-performance computing simulations on structured grids. Our approach utilizes a lossy compression scheme based on the Discrete Wavelet Transform (DWT) to achieve high compression ratios while preserving the accuracy of the simulation. Our evaluation on a finite volume scheme demonstrates that the approach can reduce memory usage by several orders of magnitude.

SOccDPT: 3D Semantic Occupancy From Dense Prediction Transformers Trained Under Memory Constraints

We present SOccDPT, a memory-efficient approach for 3D semantic occupancy prediction from monocular image input using dense prediction transformers. To address the limitations of existing methods trained on structured traffic datasets, we train our model on unstructured datasets including the Indian Driving Dataset and Bengaluru Driving Dataset. Our semisupervised training pipeline allows SOccDPT to learn from datasets with limited labels by reducing the requirement for manual labeling by substituting it with pseudo-ground truth labels to produce our Bengaluru Semantic Occupancy Dataset. This broader training enhances our model’s ability to handle unstructured traffic scenarios effectively. To overcome memory limitations during training, we introduce patch-wise training where we select a subset of parameters to train each epoch, reducing memory usage during auto-grad graph construction. In the context of unstructured traffic and memory-constrained training and inference, SOccDPT outperforms existing disparity estimation approaches as shown by the RMSE score of 9.1473, achieves a semantic segmentation IoU score of 46.02% and operates at a competitive frequency of 69.47 Hz. We make our code and semantic occupancy dataset public.

PC-mer: An Ultra-fast memory-efficient tool for metagenomics profiling and classification.

Features extraction methods, such as k-mer-based methods, have recently made up a significant role in classifying and analyzing approaches for metagenomics data. But, they are challenged by various bottlenecks, such as performance limitations, high memory consumption, and computational overhead. To deal with these challenges, we developed an innovative features extraction and sequence profiling method for DNA/RNA sequences, called PC-mer, taking advantage of the physicochemical properties of nucleotides. PC-mer in comparison with the k-mer profiling methods provides a considerable memory usage reduction by a factor of 2k while improving the metagenomics classification performance, for both machine learning-based and computational-based methods, at the various levels and also archives speedup more than 1000x for the training phase. Examining ML-based PC-mer on various datasets confirms that it can achieve 100% accuracy in classifying samples at the class, order, and family levels. Despite the k-mer-based classification methods, it also improves genus-level classification accuracy by more than 14% for shotgun dataset (i.e. achieves accuracy of 97.5%) and more than 5% for amplicon dataset (i.e. achieves accuracy of 98.6%). Due to these improvements, we provide two PC-mer-based tools, which can actually replace the popular k-mer-based tools: one for classifying and another for comparing metagenomics data.

Exploring the roles of AI-Assisted ChatGPT in the field of data science

In this study, we explore the roles of AI-assisted ChatGPT (Generative Pre-trained Transformer) in the field of data science. AI-assisted ChatGPT, a powerful language model, is fine-tuned using domain-specific data for specialised data science tasks, such as sentiment analysis and named entity recognition (NER). The results reveal significant reductions in model size and memory usage with minor trade-offs in inference time, providing valuable resource-efficient deployment. Various data augmentation methods, including back-translation, synonym replacement, and contextual word embeddings, are employed to augment the training dataset. The study's results are subjected to rigorous statistical analysis, including paired t-tests and ANOVA tests, to determine the significance of the findings. The research concludes with insightful suggestions and future scope, including advanced fine-tuning strategies, model optimization techniques, and ethical considerations.

Making Likelihood Calculations Fast: Automatic Differentiation Applied to RooFit

With the growing datasets of current and next-generation HighEnergy and Nuclear Physics (HEP/NP) experiments, statistical analysis has become more computationally demanding. These increasing demands elicit improvements and modernizations in existing statistical analysis software. One way to address these issues is to improve parameter estimation performance and numeric stability using Automatic Differentiation (AD). AD’s computational efficiency and accuracy are superior to the preexisting numerical differentiation techniques, and it offers significant performance gains when calculating the derivatives of functions with a large number of inputs, making it particularly appealing for statistical models with many parameters. For such models, many HEP/NP experiments use RooFit, a toolkit for statistical modeling and fitting that is part of ROOT. In this paper, we report on the effort to support the AD of RooFit likelihood functions. Our approach is to extend RooFit with a tool that generates overheadfree C++ code for a full likelihood function built from RooFit functional models. Gradients are then generated using Clad, a compiler-based source-codetransformation AD tool, using this C++ code. We present our results from applying AD to the entire minimization pipeline and profile likelihood calculations of several RooFit and HistFactory models at the LHC-experiment scale. We show significant reductions in calculation time and memory usage for the minimization of such likelihood functions. We also elaborate on this approach’s current limitations and explain our plans for the future.

OPTIMIZATION OF NEURAL NETWORKS FOR PREDICTING OIL RECOVERY FACTOR USING QUANTIZATION TECHNIQUES

The optimization of computational efficiency in Artificial Neural Network (ANN) models plays a cru-cial role in enhancing predictions of oil recovery factors in reservoir engineering and Enhanced Oil Re-covery (EOR). This study investigates the application of dynamic quantization to improve the efficiency of ANN models deployed in resource-constrained environments. Dynamic quantization, which converts model weights and activations to lower precision formats during inference, aims to reduce memory us-age and accelerate computation without significant loss of predictive accuracy.Using a synthetic dataset generated from the Buckley-Leverett model, encompassing parameters such as porosity, oil viscosity, permeability, classification, and time series data, we evaluated the im-pact of dynamic quantization on model size, inference time, and predictive performance. Experimental results demonstrate that dynamic quantization effectively reduces model size and speeds up inference, making it suitable for deployment on edge devices with limited computational resources.This research contributes to advancing the practical implementation of dynamic quantization tech-niques in optimizing ANN models for complex predictive tasks in reservoir engineering and related fields. The findings underscore the potential of dynamic quantization in improving computational ef-ficiency and facilitating the deployment of ANN models in real-world applications.

Bare‐Bones particle Swarm optimization‐based quantization for fast and energy efficient convolutional neural networks

AbstractNeural network quantization is a critical method for reducing memory usage and computational complexity in deep learning models, making them more suitable for deployment on resource‐constrained devices. In this article, we propose a method called BBPSO‐Quantizer, which utilizes an enhanced Bare‐Bones Particle Swarm Optimization algorithm, to address the challenging problem of mixed precision quantization of convolutional neural networks (CNNs). Our proposed algorithm leverages a new population initialization, a robust screening process, and a local search strategy to improve the search performance and guide the population towards a feasible region. Additionally, Deb's constraint handling method is incorporated to ensure that the optimized solutions satisfy the functional constraints. The effectiveness of our BBPSO‐Quantizer is evaluated on various state‐of‐the‐art CNN architectures, including VGG, DenseNet, ResNet, and MobileNetV2, using CIFAR‐10, CIFAR‐100, and Tiny ImageNet datasets. Comparative results demonstrate that our method delivers an excellent tradeoff between accuracy and computational efficiency.

Real-Time Low-Light Imaging in Space Based on the Fusion of Spatial and Frequency Domains

Due to the low photon count in space imaging and the performance bottlenecks of edge computing devices, there is a need for a practical low-light imaging solution that maintains satisfactory recovery while offering lower network latency, reduced memory usage, fewer model parameters, and fewer operation counts. Therefore, we propose a real-time deep learning framework for low-light imaging. Leveraging the parallel processing capabilities of the hardware, we perform the parallel processing of the image data from the original sensor across branches with different dimensionalities. The high-dimensional branch conducts high-dimensional feature learning in the spatial domain, while the mid-dimensional and low-dimensional branches perform pixel-level and global feature learning through the fusion of the spatial and frequency domains. This approach ensures a lightweight network model while significantly improving the quality and speed of image recovery. To adaptively adjust the image based on brightness and avoid the loss of detailed pixel feature information, we introduce an adaptive balancing module, thereby greatly enhancing the effectiveness of the model. Finally, through validation on the SID dataset and our own low-light satellite dataset, we demonstrate that this method can significantly improve image recovery speed while ensuring image recovery quality.

Fourier ptychographic microscopy image stack reconstruction using implicit neural representations

Image stacks provide invaluable 3D information in various biological and pathological imaging applications. Fourier ptychographic microscopy (FPM) enables reconstructing high-resolution, wide field-of-view image stacks without z-stack scanning, thus significantly accelerating image acquisition. However, existing FPM methods take tens of minutes to reconstruct and gigabytes of memory to store a high-resolution volumetric scene, impeding fast gigapixel-scale remote digital pathology. While deep learning approaches have been explored to address this challenge, existing methods poorly generalize to novel datasets and can produce unreliable hallucinations. This work presents FPM-INR, a compact and efficient framework that integrates physics-based optical models with implicit neural representations (INRs) to represent and reconstruct FPM image stacks. FPM-INR is agnostic to system design or sample types and does not require external training data. In our experiments, FPM-INR substantially outperforms traditional FPM algorithms with up to a 25-fold increase in speed and an 80-fold reduction in memory usage for continuous image stack representations.

VxH: A Systematic Determination of Efficient Hierarchical Voxel Structures

Three-dimensional (3D) maps with many millions to billions of points are now used in an increasing number of applications, with processing rates in the hundreds of thousands to millions of points per second. In mobile applications, power and energy consumption for managing such data and extracting useful information thereof are critical concerns. We have developed structures and methodologies with the purpose of minimizing memory usage and associated energy consumption for indexing and serialization of voxelized point-clouds. The primary source of points in our case is airborne laser scanning, but our methodology is not restricted to only this setting. Our emulated results show a memory usage reduction factor of roughly up to 200× that of Octree/Octomap and a file size reduction factor of up to 1.65× compared with the predominating compression scheme for airborne Lidar data, LASzip. In addition, our structures enable significantly more efficient processing since they are included in a hierarchical structure that captures geometric aspects.

Sustainable and lightweight domain-based intrusion detection system for in-vehicle network

Intelligent transportation systems are designed to enhance and optimize the traffic flow, safety of urban mobility, and improve energy efficiency. While advanced vehicles are equipped with new features, such as communication technologies for the exchange of safety messages, the communication interfaces of the vehicles increase the attack surfaces for attackers to exploit, even into the in-vehicle network (IVN). The automotive Ethernet and intrusion detection systems (IDSs) are promising solutions to the security problem inside the IVN. The automotive Ethernet has higher bandwidth capacity at an economical cost and flexibility for expansion than the current solution. The IDSs can protect the IVN from attackers compromising the vehicle. They can be implemented based on machine learning to learn from the normal IVN traffic behavior. However, numerous IDSs that utilize machine learning face hardware limitations when training within the in-vehicle network. Thus, the models have to be trained outside the IVN and then imported into it. Moreover, the IVN messages are unlabeled whether they are normal or under attack. We propose a lightweight unsupervised IDS that enables training in the IVN with limited computation resources. Our IDS models have an impressive parameter reduction of up to 94% compared to existing models. This leads to a remarkable reduction in memory usage of up to 86%, training time slashed by up to 69%, and a remarkable drop in energy consumption by up to 68%. Despite the size reduction, the proposed models remain only slightly less accurate than current solutions by up to 2%.

Three-dimensional global instability analysis for high-speed boundary layer flow

Herein, a numerical approach for the three-dimensional global instability analysis of high-speed boundary layer is developed. The approach is based on the Krylov–Schur method for large-scale eigenproblems. A shift-and-invert spectral transformation is employed to accelerate the convergence for eigenvalues of interest. Iterative linear solver is adopted in the transformation to reduce memory usage by avoiding the fill-in problem of direct solver and the explicit construction of matrices. A Robin condition based on BiGlobal instability analysis is imposed at the inflow boundary, and a non-reflecting condition derived from inviscid flow is used for the far-field boundary. The approach is validated in a Blasius boundary layer. According to a comparison between the performances of several algorithms in this case, the conjugate gradient squared method with no preconditioning is chosen as the linear solver. The results provided by three-dimensional global instability analysis consist with those from linear stability theory and BiGlobal instability analysis. A further demonstration of our approach is provided for a three-dimensional roughness-disturbed boundary layer. The fluctuations extracted from direct numerical simulation results illustrate that the symmetric component is the dominant unstable mode in roughness wake, as it grows twice as fast as the antisymmetric component. Two eigenvalues are obtained from the three-dimensional global instability analysis, and these represent the symmetric and antisymmetric unstable modes. The wavelengths and growth rates of both modes delivered by instability analysis agree well with the direct numerical simulation results, and the profiles of disturbance amplitude match the features of the varicose and sinuous modes.

Ensuring Data Integrity And Security In Diverse Cloud Environments To Prevent Duplicacy.

Data deduplication is a valuable technique for compressing and minimizing data duplication during data transfers, especially in cloud environments. By eliminating redundant data, it optimizes transmission capacity and reduces memory usage. To ensure the integrity of sensitive data, encryption is applied throughout the deduplication process. The SHA algorithm is commonly used for storing text data during deduplication. It generates security bits by padding the text and computes a hash consisting of hexadecimal, string, and integer data. Hash-based deduplication involves hashing the entire file and treating the hash values of text data as unique identifiers. This allows clients to identify duplicate data within the cloud. The Proof-of-Work (PoW) algorithm is widely utilized in blockchain networks like Bitcoin and Ethereum. Its primary function is to verify transactions through a process called mining. Miners engage in competition to solve complex mathematical problems, and the first one to find a solution is granted the right to add a new block to the blockchain. PoW relies on cryptographic hashing, such as the SHA-256 hashing function, to validate and secure transactions. On the other hand, the Proof of Retrievability (PoR) algorithm finds application in cloud computing systems. It serves as a consensus mechanism, ensuring that cloud storage providers store and retrieve data accurately. PoR enables cloud providers to demonstrate to consumers that their files can be fully recovered. This algorithm incorporates cryptographic proofs that validate the integrity and availability of the stored data. In cloud storage, deduplication is often implemented using the Memory mapping technique (MPT). This allows multiple data owners to store the same data in a single copy, enhancing storage efficiency. To maintain data security, encryption is applied both before and during the deduplication process, ensuring that sensitive information remains protected. In summary, data deduplication is a powerful method for compressing data, minimizing duplication, and optimizing transmission capacity. With encryption techniques and hash-based identification, it provides secure and efficient deduplication in cloud environments, while memory deduplication and MPT support further enhance performance and storage efficiency.

Bird's-Eye-View Semantic Segmentation With Two-Stream Compact Depth Transformation and Feature Rectification

Bird's-eye-view (BEV) perception has gained popularity since it provides a 3D world representation with scale consistency. Although existing camera-based solutions achieve excellent performance, the BEV positions related to features are still less accurate. In this paper, a BEV semantic segmentation framework with two-stream compact depth transformation and feature rectification is proposed. To balance the conflict that the feature maps ensemble tends to use two temporal frames with long interval, while shorter temporal frames are more beneficial to depth prediction, a two-stream compact depth transformation is designed. Between original temporal frames, we introduce an intermediate frame to decouple the joint depth estimation of original frames. The local representations of the intermediate frame are respectively matched with each original temporal frame to achieve stereo depth predictions, where compact cost volumes are built to significantly reduce memory usage with high discriminability in depth-dimension. Further, virtual camera intrinsic parameters are derived to realize adaption of compact cost volume to various 2D data augmentation and improve generalization. On this basis, BEV feature maps are obtained via feature transformation. With the influence of depth distribution errors to BEV feature map, a feature rectified segmentation network is proposed to dynamically adjust the position offsets of input features via deformable convolution and semantic information-guided feature learning. As a result, a dense and accurate BEV semantic map is obtained. In addition, a self-supervised depth estimation teacher is adopted to provide extra supervision for depth prediction of our segmentation framework. The effectiveness of the proposed method is verified on public datasets.

A Finite State Automaton for Green Data Validation in a Real-World Smart Manufacturing Environment with Special Regard to Time-Outs and Overtaking

Since data are the gold of modern business, companies put a huge effort into collecting internal and external information, such as process, supply chain, or customer data. To leverage the full potential of gathered information, data have to be free of errors and corruptions. Thus, the impacts of data quality and data validation approaches become more and more relevant. At the same time, the impact of information and communication technologies has been increasing for several years. This leads to increasing energy consumption and the associated emission of climate-damaging gases such as carbon dioxide (CO2). Since these gases cause serious problems (e.g., climate change) and lead to climate targets not being met, it is a major goal for companies to become climate neutral. Our work focuses on quality aspects in smart manufacturing lines and presents a finite automaton to validate an incoming stream of manufacturing data. Through this process, we aim to achieve a sustainable use of manufacturing resources. In the course of this work, we aim to investigate possibilities to implement data validation in resource-saving ways. Our automaton enables the detection of errors in a continuous data stream and reports discrepancies directly. By making inconsistencies visible and annotating affected data sets, we are able to increase the overall data quality. Further, we build up a fast feedback loop, allowing us to quickly intervene and remove sources of interference. Through this fast feedback, we expect a lower consumption of material resources on the one hand because we can intervene in case of error and optimize our processes. On the other hand, our automaton decreases the immaterial resources needed, such as the required energy consumption for data validation, due to more efficient validation steps. We achieve the more efficient validation steps by the already-mentioned automaton structure. Furthermore, we reduce the response time through additional recognition of overtaking data records. In addition, we implement an improved check for complex inconsistencies. Our experimental results show that we are able to significantly reduce memory usage and thus decrease the energy consumption for our data validation task.

Resultant Normal Contact Force-Based Contact Friction Model for the Combined Finite-Discrete Element Method and Its Validation

In the conventional FDEM (Combined Finite and Discrete Element Method), each contact pair might have multiple contact points where friction forces are applied, leading to non-unique friction force assignments and potentially introducing computational errors. This study introduces a new contact friction algorithm for FDEM based on the resultant normal contact force. This method necessitates determining the friction force at a unique equivalent contact point, thereby significantly simplifying the computational flow and reducing memory usage. A series of numerical tests are performed to validate the effectiveness of the proposed contact model. Using collision and block sliding tests, the proposed contact friction model is verified to be able to accurately capture the frictional effect between discrete bodies and circumvent the problematic kinetic energy dissipation issue associated with the original contact friction algorithm. For the Brazilian splitting and uniaxial compression tests, the simulated results closely align with those generated using the original contact friction algorithm and match the experimental measurements well, demonstrating the applicability of the proposed algorithm in fracturing analysis. Furthermore, by using the proposed contact friction algorithm, a computational efficiency enhancement of 8% in contact force evaluation can be achieved.

Divraster: an R package to calculate taxonomic, functional and phylogenetic diversity from rasters

Diversity metrics are widely used to better understand biodiversity patterns and help with species conservation. Taxonomic diversity alone may lack valuable information on community ecology, such as the importance of species to ecosystem functioning and evolutionary history. By combining taxonomic, functional, and phylogenetic diversity, we can complement our knowledge, particularly under climate change, by improving conservation strategies and helping to maintain biodiversity while reducing costs in the decision‐making process. However, the main tools used for computing these metrics may demand significant processing power because of their reliance on matrices, ultimately restricting their applicability. Here, we present divraster, an R package (www.r‐project.org) to calculate diversity metrics directly from species rasters, with the aim of reducing memory usage and processing power, especially when working with large datasets such as large areas, high resolution, or a high number of species. The main function of divraster can calculate changes in a given community between distinct temporal scenarios, such as present and future climate scenarios, working with taxonomic, functional, and phylogenetic diversity. It includes the partition of temporal beta diversity, making it particularly useful in ecological niche models and other macroecology studies. Additionally, divraster performs spatial calculations for both alpha and beta diversity, which are helpful in studies where time is not an issue. We also conducted a performance comparison with packages that provide similar functions and demonstrated that the divraster package outperforms them in most cases regarding memory allocation and processing time.

Scellpam: an R package/C++ library to perform parallel partitioning around medoids on scRNAseq data sets

Background Partitioning around medoids (PAM) is one of the most widely used and successful clustering method in many fields. One of its key advantages is that it only requires a distance or a dissimilarity between the individuals, and the fact that cluster centers are actual points in the data set means they can be taken as reliable representatives of their classes. However, its wider application is hampered by the large amount of memory needed to store the distance matrix (quadratic on the number of individuals) and also by the high computational cost of computing such distance matrix and, less importantly, by the cost of the clustering algorithm itself.Results Therefore, new software has been provided that addresses these issues. This software, provided under GPL license and usable as either an R package or a C++ library, calculates in parallel the distance matrix for different distances/dissimilarities (L_1, L_2, Pearson, cosine and weighted Euclidean) and also implements a parallel fast version of PAM (FASTPAM1) using any data type to reduce memory usage. Moreover, the parallel implementation uses all the cores available in modern computers which greatly reduces the execution time. Besides its general application, the software is especially useful for processing data of single cell experiments. It has been tested in problems including clustering of single cell experiments with up to 289,000 cells with the expression of about 29,000 genes per cell.Conclusions Comparisons with other current packages in terms of execution time have been made. The method greatly outperforms the available R packages for distance matrix calculation and also improves the packages that implement the PAM itself. The software is available as an R package at https://CRAN.R-project.org/package=scellpam and as C++ libraries at https://github.com/JdMDE/jmatlib and https://github.com/JdMDE/ppamlib The package is useful for single cell RNA-seq studies but it is also applicable in other contexts where clustering of large data sets is required.