Usage Of Computing Resources Research Articles

Optimizing Plant Disease Detection Models for RISC Architectures with Mini TensorFlow

Global agricultural productivity is a challenge to plant diseases, generating strong demand for rapid and efficient diagnostic tools for early detection and effective management. For plant disease classification, this study evaluates four deep learning models (EfficientNet B3, GoogLeNet, DenseNet, and VGG16). The models are evaluated based on accuracy, precision, inference latency and computational resource usage on a large dataset covering several crops and disease types while consuming CPU and RAM. However, it's found that EfficientNet B3 has better accuracy and computational efficiency compared with the other models, especially in a resource constrained setting. The rest are DenseNet, GoogLeNet and VGG16, both perform well and in DenseNet, it performs well in both accuracy and resource spent too. VGG16 only gives slightly lower accuracy than ResNet50, but it needs more computational resources. The study shows that EfficientNet B3 is a candidate for real time precision agriculture applications. The findings presented in this research offer valuable guidance about model selection for plant disease detection, enabling the development of scalable, low-cost diagnostic systems that can be used by farmers to reduce crop losses and improve yield quality.

ZeROf-Offload: forward-gradient scheme for efficient full parameter fine-tuning of billion-scale language models

Abstract In large language models (LLMs), full-parameter fine-tuning is crucial for task-specific adaptation. Traditionally, this relies on deep learning training frameworks utilizing the back-propagation scheme. However, this scheme presents inherent issues, e.g. activation memory bottlenecks and backward locking, which limit the efficient computational resource usage. In this work, we propose the design and analysis of ZeROf-Offload, an innovative fine-tuning framework that adapts the forward-gradient scheme. This framework adopts a unique forward-gradient-oriented CPU offload strategy, enabling fine-tuning of billion-scale LLMs solely in the forward phase and enhancing computational efficiency. Empirical evaluations reveal the advantage of eliminating the backward phase in fine-tuning. ZeROf-Offload achieves134 TFlops/GPU for models with over 130 billion parameters on a single DGX-A100 node, outperforming DeepSpeed’s ZeRO-Offload, which achieves 102 TFlops/GPU for models with up to 53.7 billion parameters, the largest size manageable within GPU memory limitations. Furthermore, we have expanded ZeROf-Offload for multi-DGX-A100 environments with integrated 3D parallelism, achieving near-linear speedup across up to 128 GPUs and the token throughput by 1.4x and 1.5x, respectively. The experimental results demonstrate that the proposed ZeROf-Offload has achieved the highest throughput performance compared to all examined state-of-the-art frameworks.

Dynamic Matrix Multiplication Using Reinforcement Learning

Efficient matrix multiplication is a core computational task in machine learning, deep learning, and scientific computing, as it significantly impacts the performance of model training and inference. This project explores the use of reinforcement learning (RL) to dynamically select the optimal matrix multiplication algorithm based on input matrix characteristics. We compare traditional matrix multiplication methods, including Strassen’s algorithm, Winograd’s algorithm, and tiled matrix multiplication, and implement a Q- learning-based agent to choose the most efficient algorithm adaptively. Our RL-driven approach enables real-time algorithm selection, balancing computational speed and resource usage across different matrix types and sizes. Experimental results demonstrate the system's ability to reduce execution time by selecting algorithms suited to specific scenarios, providing insights into algorithm efficiency and adaptability for various data sizes. Furthermore, this research suggests practical optimization strategies for handling large-scale computations in machine learning applications, including using parallel processing, GPU acceleration, and parameter tuning to minimize computational bottlenecks. This project underscores the potential of integrating RL for adaptive computation within AI workflows, offering a scalable solution for enhancing matrix operation efficiency in data-intensive applications

Application of Stable Diffusion and LoRA Models in AI Drawing

Abstract. This paper explores the application of Stable Diffusion model and LoRA (Low-Rank Adaptation) model in AI-generated artwork. The authors introduce the foundational principles of Stable Diffusion model and LoRA, as well as their application in high-quality image generation. Using three popular datasets ImageNet, COCO, and CelebA we apply various image quality assessment metrics, including PSNR (Peak Signal-to-Noise Ratio), IS (Inception Score), and FID (Frchet Inception Distance), and further validate their potential in artistic creation through subjective evaluations. By comparing the performance of these two models across different datasets, we examine their strengths, weaknesses, areas for improvement in image generation tasks, along with user experience considerations. The experimental results show that the Stable Diffusion model excels in terms of image quality and diversity, while the LoRA model offers significant benefits in computational efficiency and resource usage. Through comprehensive experimental evaluations, this paper provides a scientific basis for model selection in AI art creation and offers insights for the future development of hybrid models.

Method for rail surface defect detection based on neural network architecture search

Abstract This study addresses the inherent limitations of implementing neural network architecture search algorithms for rail surface defect detection, including low search efficiency and the oversight of edge features on the rail surface. A sophisticated multi-level neural network architecture search framework is proposed that integrates and emphasizes rail surface edge features. The framework utilizes the Z-Score normalization method to quantify the edge concern of rail surface defect samples, combined with an Edge-Loss function to enhance edge feature recognition capabilities. Furthermore, acknowledging the sensitivity of defect features to spatial resolution changes, a multi-level neural network architecture search space is meticulously designed. In the cell-level search space, a method combining partial channel sampling with operation pruning is employed to enhance model search efficiency and regularization. In the network-level search space, optimal paths for resolution change are established, allowing for the screening and aggregation of defect features at various levels to facilitate the adaptive extraction of multi-scale edge defect features. Experimental outcomes indicate that this method significantly reduces computational resource usage by approximately 75% and increases mIOU by 2.6% relative to traditional architecture search methods. Moreover, it demonstrates robust capability in accurately recognizing defective edges on rail surfaces, thereby substantiating the method’s effectiveness.

ML-based calibration and control of the GlueX Central Drift Chamber

The GlueX Central Drift Chamber (CDC) in Hall D at Jefferson Lab, used for detecting and tracking charged particles, is calibrated and controlled during data taking using a Gaussian process. The system dynamically adjusts the high voltage applied to the anode wires inside the chamber in response to changing environmental and experimental conditions such that the gain is stabilized. Control policies have been established to manage the CDC's behavior. These policies are activated when the model's uncertainty exceeds a configurable threshold or during human-initiated tests during normal production running. We demonstrate the system reduces the time detector experts dedicate to calibration of the data offline, leading to a marked decrease in computing resource usage without compromising detector performance.

UniSchedApi: A comprehensive solution for university resource scheduling and methodology comparison

This paper introduces UniSchedApi, an API-based solution that revolutionizes optimized university resource scheduling. The primary focus of the research is the detailed evaluation of two automatic resource allocation methods: Tabu Search (TS) and Genetic Algorithm (GA). The paper thoroughly explores how these methods address challenges associated with resource allocation in university environments, considering critical factors such as teacher availability, student time constraints, classroom features (including computers, projectors, TV's, specialized laboratories, specialized equipment, etc.), among others. The evaluation is carried out meticulously, measuring the performance and memory resource usage of both algorithms, considering the comparison with the manual scheduling. The results reveal that the TS algorithm excels in terms of temporal efficiency and computational resource usage. Based on these findings, UniSchedApi implements GA and TS but uses TS as the default algorithm, ensuring more efficient and optimized management of academic resources. This research not only presents a practical solution with UniSchedApi but also provides a deep understanding of the methods for evaluating and selecting algorithms to address specific challenges in university resource allocation. These results lay the groundwork for future improvements in academic resource management.

MiniSNV: accurate and fast single nucleotide variant calling from nanopore sequencing data.

Nanopore sequence technology has demonstrated a longer read length and enabled to potentially address the limitations of short-read sequencing including long-range haplotype phasing and accurate variant calling. However, there is still room for improvement in terms of the performance of single nucleotide variant (SNV) identification and computing resource usage for the state-of-the-art approaches. In this work, we introduce miniSNV, a lightweight SNV calling algorithm that simultaneously achieves high performance and yield. miniSNV utilizes known common variants in populations as variation backgrounds and leverages read pileup, read-based phasing, and consensus generation to identify and genotype SNVs for Oxford Nanopore Technologies (ONT) long reads. Benchmarks on real and simulated ONT data under various error profiles demonstrate that miniSNV has superior sensitivity and comparable accuracy on SNV detection and runs faster with outstanding scalability and lower memory than most state-of-the-art variant callers. miniSNV is available from https://github.com/CuiMiao-HIT/miniSNV.

Proposal of a Cost-Effective and Adaptive Customized Driver Inattention Detection Model Using Time Series Analysis and Computer Vision

Advanced Driver Assistance Systems, such as Forward Collision Warning and Lane Departure Warning, play a crucial role in accident prevention by alerting drivers to potential hazards. With the advent of fully autonomous driving technology that requires no driver input, there is now a greater emphasis on monitoring the state of vehicle occupants. This is particularly important because, in emergency situations where control must suddenly be transferred to an unprepared occupant, the risk of accidents increases significantly. To mitigate this risk, new monitoring technologies are being developed to analyze driver behavior and detect states of inattention or drowsiness. In response to the emerging demands of driver monitoring technology, we have developed the Customized Driver Inattention Detection Model (CDIDM). This model employs video analysis and statistical techniques to accurately and rapidly classify information on drivers’ gazes. The CDIDM framework defines the components of inattentive or drowsy driving based on the Driver Monitoring System (DMS) safety standards set by the European New Car Assessment Programme (EuroNCAP). By defining six driving behavior-related scenarios, we have improved the accuracy of driver inattention assessment. The CDIDM estimates the driver’s gaze while simultaneously analyzing data in real-time. To minimize computational resource usage, this model incorporates a series of preprocessing steps that facilitate efficient time series data analysis, utilizing techniques such as DTW Barycenter Averaging (DBA) and K-means clustering. This results in a robust driver attention monitoring model based on time series classification.

Enhanced you only look once approach for automatic phytoplankton identification

<p>Conventionally, identifying phytoplankton species is challenging due to human taxonomical knowledge limitations. Advanced technology can overcome this problem. A novel model that accurately enhances phytoplankton detection and identification classification by combining asymmetric convolution and vision transformers (ACVIT) within the YOLOv8m framework is promoted with ACVIT-YOLO. The performance of this model surpasses the original YOLOv8m model, exhibiting a notable 2.4% enhancement in precision, 5.5% improvement in recall, and 1.1% gain in mAP 50 score. The enhanced effectiveness of ACVIT-YOLO compared to the YOLOv8m model, further demonstrated by the decreased giga floating-point operations (GFLOP), decreased parameter count, and compact dimensions, significantly improves the automation of phytoplankton species identification. This suggests that the ACVIT-YOLO model could produce a better prediction system for identifying phytoplankton with similar accuracy to the original YOLOv8m model but with lower computational power and resource usage.</p>

Softwarized and containerized microservices-based network management analysis with MSN

Microservice architecture is a service-oriented paradigm that enables the decomposition of cumbersome monolithic-based software systems. Using microservice design principles, it is possible to develop flexible, scalable, reusable, and loosely coupled software that could be containerized and deployed in a distributed edge/cloud environment. The flexible deployment of microservices in an edge environment increases system performance in terms due to dynamic service function placement and chaining possibly resulting in latency reduction, fault tolerance, scalability, efficient resource utilization, cost reduction, and energy consumption reduction. On the other hand, virtualization and containerization of microservices add processing and communication overheads. Therefore, to evaluate end-to-end microservices-based system performance, we need to have an end-to-end mathematical formulation of the overall microservice-based network system. Incorporating the virtualization overhead, here we provide end-to-end mathematical formulation considering system parameters: latency, throughput, computational resource usage, and energy consumption. We then evaluate the formulation in a testbed environment with the Microservice-based SDN (MSN) framework that decomposes the Software-defined Networking (SDN) controller in microservices with Docker Container. The final result validates the presented mathematical modeling of the system’s dynamic behavior which can be used to design a microservice-based system.

Beyond Docker: Enhancing vantage6 with Kubernetes for Federated Learning.

Vantage6, a powerful platform for privacy-preserving analysis in the life-sciences domain, employs Docker in its data nodes when performing tasks like federated learning. This tight bond to Docker poses challenges on infrastructure security, efficient computing resources usage, and integration of alternative container technologies. To overcome these challenges, we explored the integration of Kubernetes into vantage6 through a Proof-of-Concept (PoC) approach. The PoC designed and implemented a Kubernetes-based architecture for vantage6, which can be easily deployed on a single machine or a cluster. This PoC serves as the reference to accelerate the adoption of Kubernetes in vantage6.

Enhanced Dynamic Programming Approaches for Efficient Solutions to the Traveling Salesman Problem

This study aims to enhance dynamic programming techniques for efficiently solving the Traveling Salesman Problem, a fundamental combinatorial optimization challenge. Given its NP-hard classification, traditional exact algorithms become computationally infeasible as the problem size increases. The research revisits foundational dynamic programming principles, notably the Held-Karp algorithm, and identifies existing limitations. The study begins with a comprehensive literature review, followed by an analysis of the dynamic programming challenges specific to TSP. Novel algorithms are then developed, implemented, and rigorously tested against benchmark instances. Performance evaluation is conducted using metrics such as execution time, memory usage, and solution optimality across different problem sizes. Results demonstrate significant improvements in efficiency and scalability, with enhanced algorithms achieving optimal solutions in reduced time and computational resource usage. However, the exponential growth in complexity remains a challenge for larger instances. The study concludes with recommendations for future research, focusing on further algorithmic refinements and exploring hybrid approaches to address large-scale TSPs.

Fault-tolerant quantum chemical calculations with improved machine-learning models.

Easy and effective usage of computational resources is crucial for scientific calculations. Following our recent work of machine-learning (ML) assisted scheduling optimization [J. Comput. Chem. 2023, 44, 1174], we further propose (1) the improved ML models for the better predictions of computational loads, and as such, more elaborate load-balancing calculations can be expected; (2) the idea of coded computation, that is, the integration of gradient coding, in order to introduce fault tolerance during the distributed calculations; and (3) their applications together with re-normalized exciton model with time-dependent density functional theory (REM-TDDFT) for calculating the excited states. Illustrated benchmark calculations include P38 protein, and solvent model with one or several excitable centers. The results show that the improved ML-assisted coded calculations can further improve the load-balancing and cluster utilization, owing primarily profit in fault tolerance that aims at the automated quantum chemical calculations for both ground and excited states.

SLIVER: Unveiling large scale gene regulatory networks of single-cell transcriptomic data through causal structure learning and modules aggregation

Prevalent Gene Regulatory Network (GRN) construction methods rely on generalized correlation analysis. However, in biological systems, regulation is essentially a causal relationship that cannot be adequately captured solely through correlation. Therefore, it is more reasonable to infer GRNs from a causal perspective. Existing causal discovery algorithms typically rely on Directed Acyclic Graphs (DAGs) to model causal relationships, but it often requires traversing the entire network, which result in computational demands skyrocketing as the number of nodes grows and make causal discovery algorithms only suitable for small networks with one or two hundred nodes or fewer. In this study, we propose the SLIVER (cauSaL dIscovery Via dimEnsionality Reduction) algorithm which integrates causal structural equation model and graph decomposition. SLIVER introduces a set of factor nodes, serving as abstractions of different functional modules to integrate the regulatory relationships between genes based on their respective functions or pathways, thus reducing the GRN to the product of two low-dimensional matrices. Subsequently, we employ the structural causal model (SCM) to learn the GRN within the gene node space, enforce the DAG constraint in the low-dimensional space, and guide each factor to aggregate various functions through cosine similarity. We evaluate the performance of the SLIVER algorithm on 12 real single cell transcriptomic datasets, and demonstrate it outperforms other 12 widely used methods both in GRN inference performance and computational resource usage. The analysis of the gene information integrated by factor nodes also demonstrate the biological explanation of factor nodes in GRNs. We apply it to scRNA-seq of Type 2 diabetes mellitus to capture the transcriptional regulatory structural changes of β cells under high insulin demand.

Evaluating text classification: A benchmark study

This paper presents an impartial and extensive benchmark for text classification involving five different text classification tasks, 20 datasets, 11 different model architectures, and 42,800 algorithm runs. The five text classification tasks are fake news classification, topic detection, emotion detection, polarity detection, and sarcasm detection. While in practice, especially in Natural Language Processing (NLP), research tends to focus on the most sophisticated models, we hypothesize that this is not always necessary. Therefore, our main objective is to investigate whether the largest state-of-the-art (SOTA) models are always preferred, or in what cases simple methods can compete with complex models, i.e. for which dataset specifications and classification tasks. We assess the performance of different methods with varying complexity, ranging from simple statistical and machine learning methods to pretrained transformers like robustly optimized BERT (Bidirectional Encoder Representations from Transformers) pretraining approach (RoBERTa). This comprehensive benchmark is lacking in existing literature, with research mainly comparing similar types of methods. Furthermore, with increasing awareness of the ecological impacts of extensive computational resource usage, this comparison is both critical and timely.We find that overall, bidirectional long short-term memory (LSTM) networks are ranked as the best-performing method albeit not statistically significantly better than logistic regression and RoBERTa. Overall, we cannot conclude that simple methods perform worse although this depends mainly on the classification task. Concretely, we find that for fake news classification and topic detection, simple techniques are the best-ranked models and consequently, it is not necessary to train complicated neural network architectures for these classification tasks. Moreover, we also find a negative correlation between F1 performance and complexity for the smallest datasets (with dataset size less than 10,000). Finally, the different models’ results are analyzed in depth to explain the model decisions, which is an increasing requirement in the field of text classification.

Comparative Study of Keras CNNs for Tuberculosis Detection from Chest X-rays

As a major global health burden, tuberculosis (TB) requires prompt and accurate diagnostic solutions. The effectiveness of Convolutional Neural Networks (CNNs) for TB prediction from chest X-ray images is investigated in this work using a variety of Keras applications. We compare the accuracy, efficiency, and computational resource usage of the VGG16, ResNet50, and EfficientNetV2B2 architectures. By means of comprehensive testing and analysis on a wide range of datasets, we determine the applicability of every model for tuberculosis identification. The findings show that EfficientNetV2B2 is the most promising architecture, achieving remarkable accuracy (99.5%) with enhanced computational efficiency, while VGG16 and ResNet50 show competitive performance. These results highlight how deep learning-based methods have the potential to transform tuberculosis diagnosis by providing doctors with a trustworthy and usable instrument for early detection and treatment. Our research has implications for improving healthcare outcomes and bolstering international efforts to fight tuberculosis. Keywords—EfficientNet, CNN, VGG16,Keras, Accuracy

Multi-Task Scenario Encrypted Traffic Classification and Parameter Analysis.

The widespread use of encrypted traffic poses challenges to network management and network security. Traditional machine learning-based methods for encrypted traffic classification no longer meet the demands of management and security. The application of deep learning technology in encrypted traffic classification significantly improves the accuracy of models. This study focuses primarily on encrypted traffic classification in the fields of network analysis and network security. To address the shortcomings of existing deep learning-based encrypted traffic classification methods in terms of computational memory consumption and interpretability, we introduce a Parameter-Efficient Fine-Tuning method for efficiently tuning the parameters of an encrypted traffic classification model. Experimentation is conducted on various classification scenarios, including Tor traffic service classification and malicious traffic classification, using multiple public datasets. Fair comparisons are made with state-of-the-art deep learning model architectures. The results indicate that the proposed method significantly reduces the scale of fine-tuning parameters and computational resource usage while achieving performance comparable to that of the existing best models. Furthermore, we interpret the learning mechanism of encrypted traffic representation in the pre-training model by analyzing the parameters and structure of the model. This comparison validates the hypothesis that the model exhibits hierarchical structure, clear organization, and distinct features.

BSEO-Semiautomatic Method for Determination of Oil Recovery with Nanofluids in Microfluidic Devices.

Microfluidic models have become essential instruments for studying enhanced oil recovery techniques through fluid and chemical injection into micromodels to observe interactions with pore structures and resident fluids. The widespread use of cost-effective lab-on-a-chip devices, known for efficient data extraction and minimal reagent usage, has driven demand for efficient data management methods crucial for high-performance data and image analyses. This article introduces a semiautomatic method for calculating oil recovery in polymeric nanofluid flooding experiments based on the background subtraction (BSEO). It employs the background subtraction technique, generating a foreground binary mask to detect injected fluids represented as pixel areas. The pixel difference is then compared to a threshold value to determine whether the given pixel is foreground or background. Moreover, the proposed method compares its performance with two other representative methods: the ground truth (manual segmentation) and Fiji-ImageJ software. The experiments yielded promising results. Low values of mean-squared error (MSE), mean absolute error (MAE), and root-mean-squared error (RMSE) indicate minimal prediction errors, while a substantial coefficient of determination (R2) of 98% highlights the strong correlation between the method's predictions and the observed outcomes. In conclusion, the presented method emphasizes the viability of BSEO as a robust alternative, offering the advantages of reduced computational resource usage and faster processing times.

Deep Reinforcement Learning-Empowered Cost-Effective Federated Video Surveillance Management Framework.

Video surveillance systems are integral to bolstering safety and security across multiple settings. With the advent of deep learning (DL), a specialization within machine learning (ML), these systems have been significantly augmented to facilitate DL-based video surveillance services with notable precision. Nevertheless, DL-based video surveillance services, which necessitate the tracking of object movement and motion tracking (e.g., to identify unusual object behaviors), can demand a significant portion of computational and memory resources. This includes utilizing GPU computing power for model inference and allocating GPU memory for model loading. To tackle the computational demands inherent in DL-based video surveillance, this study introduces a novel video surveillance management system designed to optimize operational efficiency. At its core, the system is built on a two-tiered edge computing architecture (i.e., client and server through socket transmission). In this architecture, the primary edge (i.e., client side) handles the initial processing tasks, such as object detection, and is connected via a Universal Serial Bus (USB) cable to the Closed-Circuit Television (CCTV) camera, directly at the source of the video feed. This immediate processing reduces the latency of data transfer by detecting objects in real time. Meanwhile, the secondary edge (i.e., server side) plays a vital role by hosting a dynamically controlling threshold module targeted at releasing DL-based models, reducing needless GPU usage. This module is a novel addition that dynamically adjusts the threshold time value required to release DL models. By dynamically optimizing this threshold, the system can effectively manage GPU usage, ensuring resources are allocated efficiently. Moreover, we utilize federated learning (FL) to streamline the training of a Long Short-Term Memory (LSTM) network for predicting imminent object appearances by amalgamating data from diverse camera sources while ensuring data privacy and optimized resource allocation. Furthermore, in contrast to the static threshold values or moving average techniques used in previous approaches for the controlling threshold module, we employ a Deep Q-Network (DQN) methodology to manage threshold values dynamically. This approach efficiently balances the trade-off between GPU memory conservation and the reloading latency of the DL model, which is enabled by incorporating LSTM-derived predictions as inputs to determine the optimal timing for releasing the DL model. The results highlight the potential of our approach to significantly improve the efficiency and effective usage of computational resources in video surveillance systems, opening the door to enhanced security in various domains.