Memory Footprint Research Articles

基于改进YOLOv8s的实时葡萄病害检测模型

This research is dedicated to enhancing the accuracy and processing speed of grape disease recognition. As a result, a real-time grape disease detection model named MSCI-YOLOv8s, based on an improved YOLOv8s framework is proposed. The primary innovation of this model lies in replacing the backbone network of the original YOLOv8s with the more efficient MobileNetV3. This alteration not only strengthens the ability of the model to capture features of various disease manifestations in grape leaf images but also improves its generalization capabilities and stability. Additionally, the model incorporates the SPPFCSPC pyramid pooling structure, which maintains the stability of the receptive field while significantly enhancing processing speed. The integration of the CBAM attention mechanism further accentuates the ability of the model to identify key features, substantially increasing the accuracy of disease detection. Moreover, the model employs Inner-SIoU as the loss function, optimizing the precision of bounding box regression and accelerating model convergence, thereby further enhancing detection efficiency. Rigorous testing has shown that the MSCI-YOLOv8s model achieves an impressive average precision (mAP) of 97.7%, with an inference time of just 37.2 milliseconds and a memory footprint of 39.3 MB. These advancements render the MSCI-YOLOv8s not only highly efficient but also extremely practical for real-time grape disease detection, meeting the actual demands of grape orchard disease identification and demonstrating significant potential for application.

DynAuthRoute: Dynamic Security for Wireless Sensor Networks

Objectives: The research aims to design an architecture for secure transmission of data in wireless sensor networks. Methods: The method involves three main pillars: authentication, data encryption, and dynamic routing. Extensive simulations have been conducted to evaluate the suggested method in terms of energy consumption, memory footprint, packet delivery ratio, end-to-end latency, execution time, encryption time, and decryption time. Findings: For authentication, a dynamic key is used to power an improved salt password hashing method. Data encryption is performed using format-preserving encryption (FPE) with the appended salt key. Dynamic routing is implemented using a cluster-based routing technique to enhance network efficiency in terms of power consumption and security. The execution time for MD5 ranges from 15 to 22 milliseconds, while for SHA-1 it ranges from 16 to 23 milliseconds and for the proposed salt key generation it is 1 to 5 milliseconds. Similarly, in terms of energy consumption, memory footprint, packet delivery ratio, end-to-end latency, execution time, encryption time, and decryption time the proposed method shows promising results in ensuring the integrity and security of transmitted encrypted data. Novelty: The presents a novel architecture with enhanced cluster head-based selection algorithm that combines dynamic key-based authentication and secure data routing to establish a safe environment for data transmission in wireless sensor networks. This research works offers a method for encrypting text with a dynamic salt key that is safe, energy-efficient, and lightweight. Keywords: Wireless Sensor Network, Dynamic Key, Authentication, Hash function, Salt algorithm, Dynamic routing, Node clustering, Format-preserving encryption

Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training

Traditional adversarial training, while effective at improving machine learning model robustness, is computationally intensive. Fast Adversarial Training (FAT) addresses this by using a single-step attack to generate adversarial examples more efficiently. Nonetheless, FAT is susceptible to a phenomenon known as catastrophic overfitting, wherein the model's adversarial robustness abruptly collapses to zero during the training phase. To address this challenge, recent studies have suggested adopting adversarial initialization with Fast Gradient Sign Method Adversarial Training (FGSM-AT), which recycles adversarial perturbations from prior epochs by computing gradient momentum. However, our research has uncovered a flaw in this approach. Given that data augmentation is employed during the training phase, the samples in each epoch are not identical. Consequently, the method essentially yields not the adversarial perturbation of a singular sample, but rather the Universal Adversarial Perturbation (UAP) of a sample and its data augmentation. This insight has led us to explore the potential of using UAPs for adversarial initialization within the context of FGSM-AT. We have devised various strategies for adversarial initialization utilizing UAPs, including single, class-based, and feature-based UAPs. Experiments conducted on three distinct datasets demonstrate that our method achieves an improved trade-off among robustness, computational cost, and memory footprint. Code is available at https://github.com/fzjcdt/fgsm-uap.

G-Adapter: Towards Structure-Aware Parameter-Efficient Transfer Learning for Graph Transformer Networks

It has become a popular paradigm to transfer the knowledge of large-scale pre-trained models to various downstream tasks via fine-tuning the entire model parameters. However, with the growth of model scale and the rising number of downstream tasks, this paradigm inevitably meets the challenges in terms of computation consumption and memory footprint issues. Recently, Parameter-Efficient Fine-Tuning (PEFT) (e.g., Adapter, LoRA, BitFit) shows a promising paradigm to alleviate these concerns by updating only a portion of parameters. Despite these PEFTs having demonstrated satisfactory performance in natural language processing, it remains under-explored for the question: whether these techniques could be transferred to graph-based tasks with Graph Transformer Networks (GTNs)? Therefore, in this paper, we fill this gap by providing extensive benchmarks with traditional PEFTs on a range of graph-based downstream tasks. Our empirical study shows that it is sub-optimal to directly transfer existing PEFTs to graph-based tasks due to the issue of feature distribution shift. To address this issue, we propose a novel structure-aware PEFT approach, named G-Adapter, which leverages graph convolution operation to introduce graph structure information (e.g., graph adjacency matrix) as an inductive bias to guide the updating process. Further, we propose Bregman proximal point optimization to alleviate feature distribution shift by preventing the model from aggressive update. Extensive experiments demonstrate that G-Adapter obtains state-of-the-art performance compared to counterparts on nine graph benchmark datasets based on diverse pre-trained GTNs, and delivers tremendous memory footprint efficiency compared to the conventional paradigm.

Deep Hierarchical Video Compression

Recently, probabilistic predictive coding that directly models the conditional distribution of latent features across successive frames for temporal redundancy removal has yielded promising results. Existing methods using a single-scale Variational AutoEncoder (VAE) must devise complex networks for conditional probability estimation in latent space, neglecting multiscale characteristics of video frames. Instead, this work proposes hierarchical probabilistic predictive coding, for which hierarchal VAEs are carefully designed to characterize multiscale latent features as a family of flexible priors and posteriors to predict the probabilities of future frames. Under such a hierarchical structure, lightweight networks are sufficient for prediction. The proposed method outperforms representative learned video compression models on common testing videos and demonstrates computational friendliness with much less memory footprint and faster encoding/decoding. Extensive experiments on adaptation to temporal patterns also indicate the better generalization of our hierarchical predictive mechanism. Furthermore, our solution is the first to enable progressive decoding that is favored in networked video applications with packet loss.

DTL: Disentangled Transfer Learning for Visual Recognition

When pre-trained models become rapidly larger, the cost of fine-tuning on downstream tasks steadily increases, too. To economically fine-tune these models, parameter-efficient transfer learning (PETL) is proposed, which only tunes a tiny subset of trainable parameters to efficiently learn quality representations. However, current PETL methods are facing the dilemma that during training the GPU memory footprint is not effectively reduced as trainable parameters. PETL will likely fail, too, if the full fine-tuning encounters the out-of-GPU-memory issue. This phenomenon happens because trainable parameters from these methods are generally entangled with the backbone, such that a lot of intermediate states have to be stored in GPU memory for gradient propagation. To alleviate this problem, we introduce Disentangled Transfer Learning (DTL), which disentangles the trainable parameters from the backbone using a lightweight Compact Side Network (CSN). By progressively extracting task-specific information with a few low-rank linear mappings and appropriately adding the information back to the backbone, CSN effectively realizes knowledge transfer in various downstream tasks. We conducted extensive experiments to validate the effectiveness of our method. The proposed method not only reduces a large amount of GPU memory usage and trainable parameters, but also outperforms existing PETL methods by a significant margin in accuracy, achieving new state-of-the-art on several standard benchmarks.

ISwap: A New Memory Page Swap Mechanism for Reducing Ineffective I/O Operations in Cloud Environments

This paper proposes iSwap, a new memory page swap mechanism that reduces the ineffective I/O swap operations and improves the QoS for applications with a high priority in the cloud environments. iSwap works in the OS kernel. iSwap accurately learns the reuse patterns for memory pages and makes the swap decisions accordingly to avoid ineffective operations. In the cases where memory pressure is high, iSwap compresses pages that belong to the latency-critical (LC) applications (or high-priority applications) and keeps them in main memory, avoiding I/O operations for these LC applications to ensure QoS; and iSwap evicts low-priority applications’ pages out of main memory. iSwap has a low overhead and works well for cloud applications with large memory footprints. We evaluate iSwap on Intel x86 and ARM platforms. The experimental results show that iSwap can significantly reduce ineffective swap operations (8.0% - 19.2%) and improve the QoS for LC applications (36.8% - 91.3%) in cases where memory pressure is high, compared with the latest LRU-based approach widely used in modern OSes.

Comparing apples and oranges? Investigating the consistency of CPU and memory profiler results across multiple java versions

Profiling is an important tool in the software developer’s box, used to identify hot methods where most computational resources are used, to focus efforts at improving efficiency. Profilers are also important in the context of Genetic improvement (GI) of software. GI applies search-based optimisation to existing software with many examples of success in a variety of contexts. GI generates variants of the original program, testing each for functionality and properties such as run time or memory footprint, and profiling can be used to target the code variations to increase the search efficiency. We report on an experimental study comparing two profilers included with different versions of the Java Development Kit (JDK), HPROF (JDK 8) and Java Flight Recorder (JFR) (JDK 8, 9, and 17), within the GI toolbox Gin on six open-source applications, for both run time and memory use. We find that a core set of methods are labelled hot in most runs, with a long tail appearing rarely. We suggest five repeats enough to overcome this noise. Perhaps unsurprisingly, changing the profiler and JDK dramatically change the hot methods identified, so profiling must be rerun for new JDKs. We also show that using profiling for test case subset selection is unwise, often missing relevant members of the test suite. Similar general patterns are seen for memory profiling as for run time but the identified hot methods are often quite different.

Encrypted Image Classification with Low Memory Footprint Using Fully Homomorphic Encryption.

Classifying images has become a straightforward and accessible task, thanks to the advent of Deep Neural Networks. Nevertheless, not much attention is given to the privacy concerns associated with sensitive data contained in images. In this study, we propose a solution to this issue by exploring an intersection between Machine Learning and cryptography. In particular, Fully Homomorphic Encryption (FHE) emerges as a promising solution, as it enables computations to be performed on encrypted data. We therefore propose a Residual Network implementation based on FHE which allows the classification of encrypted images, ensuring that only the user can see the result. We suggest a circuit which reduces the memory requirements by more than [Formula: see text] compared to the most recent works, while maintaining a high level of accuracy and a short computational time. We implement the circuit using the well-known Cheon-Kim-Kim-Song (CKKS) scheme, which enables approximate encrypted computations. We evaluate the results from three perspectives: memory requirements, computational time and calculations precision. We demonstrate that it is possible to evaluate an encrypted ResNet20 in less than five minutes on a laptop using approximately 15[Formula: see text]GB of memory, achieving an accuracy of 91.67% on the CIFAR-10 dataset, which is almost equivalent to the accuracy of the plain model (92.60%).

Assessment of inference accuracy and memory capacity of computation-in-memory enabled neural network due to quantized weights, gradients, input and output signals, and memory non-idealities

This paper proposes an approach to enhance the efficiency of computation-in-memory (CiM) enabled neural networks. The proposed methods involve partial quantization of learning and inference processes within the neural network to increase the training and inference speed while reducing energy and memory consumption. The impact of quantization due to the usage of CiM is evaluated based on inference accuracy. The effect of non-idealities incurred due to the employment of different memories such as resistive random-access memory on the network accuracy is documented and reported. The results indicate that a certain quantization bit precision threshold is necessary for weights, input/output data, and gradients to maintain an acceptable inference accuracy level. Notably, the experiments demonstrate a modest degradation of approximately 2.8% in inference accuracy compared to the neural network trained without using computation-in-memory, this accuracy trade-off is accompanied by a substantial memory footprint improvement, with memory usage reductions of 62% and 93% during the training and inference phase respectively.

Lighter and faster simulations on domains with symmetries

A strategy to improve the performance and reduce the memory footprint of simulations on meshes with spatial reflection symmetries is presented in this work. By using an appropriate mirrored ordering of the unknowns, discrete partial differential operators are represented by matrices with a regular block structure that allows replacing the standard sparse matrix–vector product with a specialised version of the sparse matrix-matrix product, which has a significantly higher arithmetic intensity. Consequently, matrix multiplications are accelerated, whereas their memory footprint is reduced, making massive simulations more affordable. As an example of practical application, we consider the numerical simulation of turbulent incompressible flows using a low-dissipation discretisation on unstructured collocated grids. All the required matrices are classified into three sparsity patterns that correspond to the discrete Laplacian, gradient, and divergence operators. Therefore, the above-mentioned benefits of exploiting spatial reflection symmetries are tested for these three matrices on both CPU and GPU, showing up to 5.0x speed-ups and 8.0x memory savings. Finally, a roofline performance analysis of the symmetry-aware sparse matrix–vector product is presented.

Development of multilevel mesh adaptivity for global variance reduction in Monte Carlo code cosRMC

The mesh-based weight window method is a widely used global variance reduction technique for Monte Carlo particle transport calculation. In order to meet the fine voxel division required for high-quality weight window distribution, the uniform structured mesh division will lead to large memory footprint pressure. This work focused on the research of the multilevel adaptive mesh algorithm, using a recursive algorithm to adjust the voxel resolution adaptively according to the material of geometry. Then the adaptive mesh algorithm was applied to the weight window generator based on the response matrix method by establishing the neighbor relationship of adaptive mesh. And an optimization method for memory allocation of the response matrix solver was proposed, effectively reducing the memory footprint of the response matrix solver in parallel computing. In the calculation of the CFETR model, the calculation efficiency under the weight window based on adaptive mesh was increased by 53.3 % compared with the weight window based on uniform mesh, and the proportion of voxels with a statistical error less than 10 % reached 93.33 %, which verified the effectiveness of the multilevel adaptive mesh algorithm.

Algorithm 1039: Automatic Generators for a Family of Matrix Multiplication Routines with Apache TVM

We explore the utilization of the Apache TVM open source framework to automatically generate a family of algorithms that follow the approach taken by popular linear algebra libraries, such as GotoBLAS2, BLIS, and OpenBLAS, to obtain high-performance blocked formulations of the general matrix multiplication ( gemm ). In addition, we fully automatize the generation process by also leveraging the Apache TVM framework to derive a complete variety of the processor-specific micro-kernels for gemm . This is in contrast with the convention in high-performance libraries, which hand-encode a single micro-kernel per architecture using Assembly code. In global, the combination of our TVM-generated blocked algorithms and micro-kernels for gemm (1) improves portability, maintainability, and, globally, streamlines the software life cycle; (2) provides high flexibility to easily tailor and optimize the solution to different data types, processor architectures, and matrix operand shapes, yielding performance on a par (or even superior for specific matrix shapes) with that of hand-tuned libraries; and (3) features a small memory footprint.

GraphScale: Scalable Processing on FPGAs for HBM and Large Graphs

Recent advances in graph processing on FPGAs promise to alleviate performance bottlenecks with irregular memory access patterns. Such bottlenecks challenge performance for a growing number of important application areas like machine learning and data analytics. While FPGAs denote a promising solution through flexible memory hierarchies and massive parallelism, we argue that current graph processing accelerators either use the off-chip memory bandwidth inefficiently or do not scale well across memory channels. In this work, we propose GraphScale, a scalable graph processing framework for FPGAs. GraphScale combines multi-channel memory with asynchronous graph processing (i.e., for fast convergence on results) and a compressed graph representation (i.e., for efficient usage of memory bandwidth and reduced memory footprint). GraphScale solves common graph problems like breadth-first search, PageRank, and weakly connected components through modular user-defined functions, a novel two-dimensional partitioning scheme, and a high-performance two-level crossbar design. Additionally, we extend GraphScale to scale to modern high-bandwidth memory (HBM) and reduce partitioning overhead of large graphs with binary packing.

Multi-Scale Residual Spectral–Spatial Attention Combined with Improved Transformer for Hyperspectral Image Classification

Aiming to solve the problems of different spectral bands and spatial pixels contributing differently to hyperspectral image (HSI) classification, and sparse connectivity restricting the convolutional neural network to a globally dependent capture, we propose a HSI classification model combined with multi-scale residual spectral–spatial attention and an improved transformer in this paper. First, in order to efficiently highlight discriminative spectral–spatial information, we propose a multi-scale residual spectral–spatial feature extraction module that preserves the multi-scale information in a two-layer cascade structure, and the spectral–spatial features are refined by residual spectral–spatial attention for the feature-learning stage. In addition, to further capture the sequential spectral relationships, we combine the advantages of Cross-Attention and Re-Attention to alleviate computational burden and attention collapse issues, and propose the Cross-Re-Attention mechanism to achieve an improved transformer, which can efficiently alleviate the heavy memory footprint and huge computational burden of the model. The experimental results show that the overall accuracy of the proposed model in this paper can reach 98.71%, 99.33%, and 99.72% for Indiana Pines, Kennedy Space Center, and XuZhou datasets, respectively. The proposed method was verified to have high accuracy and effectiveness compared to the state-of-the-art models, which shows that the concept of the hybrid architecture opens a new window for HSI classification.

SWIX: A Memory-efficient Sliding Window Learned Index

Data stream processing systems enable querying over sliding windows of streams of data. Efficient index structures for the streaming window are a crucial building block to enable querying the sliding window for operations such as aggregation and joins. This paper proposes SWIX, a novel memory-efficient learned index for sliding windows. Unlike conventional learned indexes that rely on tree structures to achieve logarithmic query cost, SWIX has a flat structure that uses substantially less memory and enables efficient query execution while having a low cost for index maintenance when inserting (and retraining). SWIX dynamically adapts itself to the real-time distribution shifts of data streams. SWIX outperforms existing indexes in terms of query execution time and memory footprint for workloads characterized by very frequent updates. Our results show that SWIX has a significantly smaller memory footprint than conventional, streaming, and learned indexes, using only 22% to 42% of the size compared to state-of-the-art approaches, yet outperforming them by up 1.2× to 1.6× on average (and up to 52×) in terms of query time, making it a space- and time-efficient method for indexing data streams. For concurrent learned indexes, Parallel SWIX can achieve up to 3.45× throughput with only 34% of memory consumption.

Spruce: a Fast yet Space-saving Structure for Dynamic Graph Storage

Dynamic graphs have been gaining increasing popularity across various application domains. With the growing size of these graphs, the update performance as well as space occupancy is becoming a crucial aspect of dynamic graph storage. Although existing dynamic graph systems can handle massive streaming updates (e.g., insertions and deletions), they cannot achieve both high throughput and low memory footprint. Drawing inspiration from the basic operations of the van Emde Boas (vEB) tree in double-logarithmic time, we designed Spruce, a high-performance yet space-saving in-memory structure to store dynamic graphs. Spruce uses a compact representation to construct the tree-like multilevel structure, which shares the common prefixes of vertices and has no merging or splitting of nodes to achieve the requirements of low memory consumption and high-efficiency dynamic operations. Furthermore, Spruce incorporates a read-optimized concurrency protocol, which refines ROWEX and Optimistic Locking, to facilitate efficient simultaneous read/write operations. Our experiment demonstrates that compared to Sortledton (the best of competitors), Spruce is up to 2.4X faster in ingesting graph updates, while saving up to 38.5% of memory space. As for graph analytics, Spruce shows high adaptability to different analytical workloads, and achieves comparable performance to other state-of-the-art dynamic graph structures.

An Efficient Method for Solving Problems of Acoustic Scattering on Three-Dimensional Transparent Structures

The article contains a study of methods for solving integral equations in the context of acoustic problems. The methodology considered is applied to describe acoustic wave propagation and scattering. Efficient discretization methods are used together with iterative methods to solve the operator equations, including an apparatus for fast multiplication of the resulting post-discretization Toeplitz matrices by a vector using the fast Fourier transform. The theoretical analysis of the proposed numerical algorithm demonstrates its efficiency in terms of the required number of arithmetic operations and the memory footprint of the computing system. The presented numerical simulation demonstrates the possibility of solving the problem of acoustic wave propagation in transparent media using the proposed methods. A visualization of the obtained solutions for a practical problem with a high level of discretization of the solution volume domain is also presented.

AutoQNN: An End-to-End Framework for Automatically Quantizing Neural Networks

Exploring the expected quantizing scheme with suitable mixed-precision policy is the key point to compress deep neural networks (DNNs) in high efficiency and accuracy. This exploration implies heavy workloads for domain experts, and an automatic compression method is needed. However, the huge search space of the automatic method introduces plenty of computing budgets that make the automatic process challenging to be applied in real scenarios. In this paper, we propose an end-to-end framework named AutoQNN, for automatically quantizing different layers utilizing different schemes and bitwidths without any human labor. AutoQNN can seek desirable quantizing schemes and mixed-precision policies for mainstream DNN models efficiently by involving three techniques: quantizing scheme search (QSS), quantizing precision learning (QPL), and quantized architecture generation (QAG). QSS introduces five quantizing schemes and defines three new schemes as a candidate set for scheme search, and then uses the differentiable neural architecture search (DNAS) algorithm to seek the layer- or model-desired scheme from the set. QPL is the first method to learn mixed-precision policies by reparameterizing the bitwidths of quantizing schemes, to the best of our knowledge. QPL optimizes both classification loss and precision loss of DNNs efficiently and obtains the relatively optimal mixed-precision model within limited model size and memory footprint. QAG is designed to convert arbitrary architectures into corresponding quantized ones without manual intervention, to facilitate end-to-end neural network quantization. We have implemented AutoQNN and integrated it into Keras. Extensive experiments demonstrate that AutoQNN can consistently outperform state-of-the-art quantization.

Increasing accuracy of pulse arrival time estimation in low frequency recordings

Objective. Wearable devices that measure vital signals using photoplethysmography are becoming more commonplace. To reduce battery consumption, computational complexity, memory footprint or transmission bandwidth, companies of commercial wearable technologies are often looking to minimize the sampling frequency of the measured vital signals. One such vital signal of interest is the pulse arrival time (PAT), which is an indicator of blood pressure. To leverage this non-invasive and non-intrusive measurement data for use in clinical decision making, the accuracy of obtained PAT-parameters needs to increase in lower sampling frequency recordings. The aim of this paper is to develop a new strategy to estimate PAT at sampling frequencies up to 25 Hertz. Approach. The method applies template matching to leverage the random nature of sampling time and expected change in the PAT. Main results. The algorithm was tested on a publicly available dataset from 22 healthy volunteers, under sitting, walking and running conditions. The method significantly reduces both the mean and the standard deviation of the error when going to lower sampling frequencies by an average of 16.6% and 20.2%, respectively. Looking only at the sitting position, this reduction is even larger, increasing to an average of 22.2% and 48.8%, respectively. Significance. This new method shows promise in allowing more accurate estimation of PAT even in lower frequency recordings.