UTSA Runner Research Press (Univ. of Texas at San Antonio)
Not a member yet
6846 research outputs found
Sort by
Realigned Softmax Warping for Compact and Separable Embedding Formation
No full textDeep Metric Learning (DML) loss functions traditionally aim to control the forces of separability and compactness within an embedding space so that the same class data points are pulled together and different class ones are pushed apart. Within the context of DML, a softmax operation will typically normalize distances into a probability for optimization, thus coupling all the push/pull forces together. This work proposes a potential new class of loss functions that operate within a euclidean domain and aim to take full advantage of the coupled forces governing embedding space formation under a softmax. These forces of compactness and separability can beboosted or mitigated within controlled locations at will by using a warping function. In this work, a simple example of a warping function is provided and then used to achieve competitive, state-of-the-art results on various Image Retrieval and Face Recognition benchmarks. Furthermore, building upon the language of warped landscapes, a new insight is also provided into the nature of cosine losses.Electrical and Computer Engineerin
Parameter Estimation of Birnbaum-Saunders Distribution under Competing Risks Using the Quantile Variant of the Expectation-Maximization Algorithm
No full textCompeting risks models, also known as weakest-link models, are utilized to analyze diverse strength distributions exhibiting multi-modality, often attributed to various types of defects within the material. The weakest-link theory posits that a material's fracture is dictated by its most severe defect. However, multimodal problems can become intricate due to potential censoring, a common constraint stemming from time and cost limitations during experiments. Additionally, determining the mode of failure can be challenging due to factors like the absence of suitable diagnostic tools, costly autopsy procedures, and other obstacles, collectively referred to as the masking problem. In this paper, we investigate the distribution of strength for multimodal failures with censored data. We consider both full and partial maskings and present an EM-type parameter estimate for the Birnbaum-Saunders distribution under competing risks. We compare the results with those obtained from other distributions, such as lognormal, Weibull, and Wald (inverse-Gaussian) distributions. The effectiveness of the proposed method is demonstrated through two illustrative examples, as well as an analysis of the sensitivity of parameter estimates to variations in starting values.Management Science and Statistic
Ensuring Resource-Efficiency and Trustworthiness in Machine Learning Systems
No full textRecent advancements in Machine Learning (ML) have significantly elevated the capabilities of artificial intelligence, driving breakthroughs across diverse domains. However, training and deploying ML models encounter critical challenges, such as the preservation of user privacy and optimization of resource utilization, especially in decentralized and resource-constrained environments. This dissertation addresses these challenges by introducing a series of innovative methods aimed at enhancing model performance, ensuring privacy preservation, and improving resource efficiency in federated learning (FL) systems and large language models (LLMs).
We first tackle resource optimization in the FL systems by employing Lyapunov optimization to guide client sampling and resource allocation. This approach minimizes overall training latency while maintaining model performance under uncertain communication environments, heterogeneous device capabilities, and energy constraints. By dynamically selecting clients based on their available resources and data characteristics, we reduce the impact of stragglers and enhance the efficiency of the FL process.
Building upon this, we develop an optimization framework for a two-tier hierarchical federated learning (HFL) system. By integrating dynamic topology design and resource management, we focus on improving training efficiency and scalability. The proposed framework optimizes both network topology and resource allocation, minimizing training latency through strategic allocation of computational and communication resources and adjustments to peer-to-peer connections among edge servers. A model consensus constraint is introduced to ensure convergence under dynamic network settings, facilitating the effective deployment of large-scale federated learning in edge networks.
In the context of large models, we address the challenges of fine-tuning large-scale generative models by introducing two parameter-efficient fine-tuning methods: LoHa and LoKr. These methods enhance the adaptability of models like Stable Diffusion without increasing the number of trainable parameters, overcoming the limitations of traditional low-rank constraints while retaining parameter efficiency. LoHa increases the maximum rank of weight update matrices, enabling better adaptation to diverse downstream tasks, while LoKr provides finer granularity in model adaptation through Kronecker product-based decomposition.
Furthermore, we develop FedPT, a novel framework for federated fine-tuning of large language models (LLMs) in resource-constrained environments. FedPT addresses the memory, computation, and communication challenges inherent in federated fine-tuning of LLMs by collaboratively training a smaller proxy model on edge devices. This proxy model captures domain-specific knowledge from distributed clients and guides the output logits of a larger model on the server. This approach significantly reduces computational and communication overhead, making the fine-tuning of LLMs feasible in federated settings without compromising performance.
Additionally, we propose FAQ-HLoRA, an adaptive framework that combines adaptive quantization and heterogeneous Low-Rank Adaptation (LoRA) ranks to address both data and system heterogeneity in federated fine-tuning. FAQ-HLoRA dynamically adjusts quantization levels and LoRA ranks based on each device's available resources, mitigating the straggler effect and improving training efficiency while maintaining model performance. By employing a per-round optimization strategy that jointly considers training delay, model accuracy, and device resource constraints, we achieve efficient large language model training across heterogeneous devices.
Through these innovations, this dissertation makes substantial contributions to developing efficient, scalable, and privacy-preserving ML systems. These advancements enable efficient model training and fine-tuning in decentralized, resource-constrained environments, facilitating real-world applications from healthcare to smart infrastructure.Electrical and Computer Engineerin
Correction: Gautam et al. Experimental Thermal Conductivity Studies of Agar-Based Aqueous Suspensions with Lignin Magnetic Nanocomposites. Magnetochemistry 2024, 10, 12
No full textIn the original publication [1] there was a unit error in Figure 5. [...]Biomedical Engineering and Chemical Engineerin
An Autochthonous Susceptible Candida auris Clade I Otomycosis Case in Iran
Get PDF<i>Candida auris</i> is a newly emerging multidrug-resistant fungal pathogen considered to be a serious global health threat. Due to diagnostic challenges, there is no precise estimate for the prevalence rate of this pathogen in Iran. Since 2019, only six culture-proven <i>C. auris</i> cases have been reported from Iran, of which, five belonged to clade V and one to clade I. Herein, we report a case of otomycosis due to <i>C. auris</i> from 2017 in a 78-year-old man with diabetes mellitus type II without an epidemiological link to other cases or travel history. Short tandem repeat genotyping and whole genome sequencing (WGS) analysis revealed that this isolate belonged to clade I of <i>C. auris</i> (South Asian Clade). The WGS single nucleotide polymorphism calling demonstrated that the <i>C. auris</i> isolate from 2017 is not related to a previously reported clade I isolate from Iran. The presence of this retrospectively recognized clade I isolate also suggests an early introduction from other regions or an autochthonous presence. Although the majority of reported <i>C. auris</i> isolates worldwide are resistant to fluconazole and, to a lesser extent, to echinocandins and amphotericin B, the reported clade I isolate from Iran was susceptible to all antifungal drugs.Molecular Microbiology & ImmunologySouth Texas Center for Emerging Infectious Disease
Leveraging a Delay Tolerant Network Approach for Critical Network Architectures
Get PDFAdvancements in network and communication technologies have enabled new network architectures to connect a growing amount of people and devices. Critical network architectures can reside in challenged environments that pose risks to network reliability and performance. Challenged environments constitute areas where assumptions of end-to-end connectivity cannot be made and as a result classical network models, such as TCP/IP, may not be suitable. In this work, a Delay Tolerant Network (DTN) framework is applied to a critical manufacturing use case scenario. The tradeoffs of using a DTN approach are evaluated in a commercial cloud environment via simulation of data pipelines running the High-rate DTN (HDTN) software implementation. Experimentation revealed inconsistent dynamic routing and performance degradation as current issues with the HDTN software that need to be addressed to make it a viable solution in critical manufacturing environments. For the dynamic routing problem, the behavior is that the router module prefers to utilize an active HDTN router for bundle processing. When reviewing performance degradation there are opportunities for improvement through tuning of convergence layer adapters or hardware acceleration, if applicable. While issues were observed in using the HDTN overlay, this work is an early attempt to incorporate HDTN in a manufacturing use case scenario. In this research, we seek to provide a baseline in exploring a DTN solution for critical manufacturing architectures for future delay tolerant applications.Electrical and Computer Engineerin
Empathy Driven Social Emotional Learning (SEL): Unraveling the Role of the Teacher Through Nexus Analysis
No full textIn the realm of elementary education, the influence of teacher empathy on Social and Emotional Learning (SEL) remains a critical yet nuanced aspect. This qualitative study explores the integral role of teacher empathy in shaping SEL environments in K-5 classrooms. Teacher empathy, defined as the capacity to perceive situations from students' perspectives, fosters a supportive and understanding learning environment. The study explores the evidence of empathy in classroom practices and its impact on teachers' understanding of their role in influencing SEL instruction, with a focus on four core competencies: self-regulation, social awareness, relationship skill-building, and responsible decision-making. The research aims to understand the impact of teacher empathy on student outcomes, contributing to a positive school climate and reduced misbehaviors. Employing a nexus analytical methodology, the study delves into the teacher's historical background, interaction order, and classroom discourses. Data analysis revealed nine themes through qualitative coding, showcasing the significant role of teacher empathy in informing pedagogy for self-regulation and social awareness. While inconclusive findings were observed for responsible decision-making and relationship skill-building, discourse analysis unveiled disparities between professed practices and observed behaviors. The areas for future growth in the study included methodological constraints, contextual factors of three participants, and applicability of findings to the real-world. Recommendations encompass comprehensive SEL training, continuous evaluation of teacher empathy, and targeted professional development to enhance emotional competencies, creating an inclusive and conducive learning environment.Bicultural-Bilingual Studie
A Computationally Time-Efficient Method for Implementing Pressure Load to FE Models with Lagrangian Elements
No full textA computationally time-efficient method is introduced to implement pressure load to a Finite element model. Hexahedron elements of the Lagrangian family with Gauss&ndash;Lobatto nodes and integration quadrature are utilized, where the integration points follow the same sequence as the nodes. This method calculates the equivalent nodal force due to pressure load using a single Hadamard multiplication. The arithmetic operations of this method are determined, which affirms its computational efficiency. Finally, the method is tested with finite element implementation and observed to increase the runtime ratio compared to the conventional method by over 20 times. This method can benefit the implementation of finite element models in fields where computational time is crucial, such as real-time and cyber&ndash;physical testbed implementation
An Ensemble Deep CNN Approach for Power Quality Disturbance Classification: A Technological Route Towards Smart Cities Using Image-Based Transfer
No full textThe abundance of powered semiconductor devices has increased with the introduction of renewable energy sources into the grid, causing power quality disturbances (PQDs). This represents a huge challenge for grid reliability and smart city infrastructures. Accurate detection and classification are important for grid reliability and consumers&rsquo; appliances in a smart city environment. Conventionally, power quality monitoring relies on trivial machine learning classifiers or signal processing methods. However, recent advancements have introduced Deep Convolution Neural Networks (DCNNs) as promising methods for the detection and classification of PQDs. These techniques have the potential to demonstrate high classification accuracy, making them a more appropriate choice for real-time operations in a smart city framework. This paper presents a voting ensemble approach to classify sixteen PQDs, using the DCNN architecture through transfer learning. In this process, continuous wavelet transform (CWT) is employed to convert one-dimensional (1-D) PQD signals into time&ndash;frequency images. Four pre-trained DCNN architectures, i.e., Residual Network-50 (ResNet-50), Visual Geometry Group-16 (VGG-16), AlexNet and SqeezeNet are trained and implemented in MATLAB, using images of four datasets, i.e., without noise, 20 dB noise, 30 dB noise and random noise. Additionally, we also tested the performance of ResNet-50 with a squeeze-and-excitation (SE) mechanism. It was observed that ResNet-50 with the SE mechanism has a better classification accuracy; however, it causes computational overheads. The classification performance is enhanced by using the voting ensemble model. The results indicate that the proposed scheme improved the accuracy (99.98%), precision (99.97%), recall (99.80%) and F1-score (99.85%). As an outcome of this work, it is demonstrated that ResNet-50 with the SE mechanism is a viable choice as a single classification model, while an ensemble approach further increases the generalized performance for PQD classification.Electrical and Computer Engineerin
Containerized Computer Vision Applications on Arm-Powered Edge Devices
Get PDFThe rise of IoT devices has led to an increased use of computer vision applications. However, the traditional IoT-Cloud model struggles to process image data efficiently due to bandwidth and latency issues. To overcome these obstacles, edge nodes have been introduced between IoT devices and the cloud. However, the deployment and securing of computer vision applications on these nodes remains a challenge. Container technology offers a promising solution, but its performance in specific domains has not been fully investigated.
In response to this challenge, this doctoral dissertation explores the effectiveness of using lightweight container technology to deploy CPU-based computer vision applications on edge devices through three primary approaches. First, it assesses various container technologies and images in computer vision applications and scrutinizes their performance across various CPUs and GPUs on ARM-based edge devices. Second, it constructs an optimized OpenCV container to serve as the foundational image for containerized computer vision applications. Lastly, the performance of the optimized OpenCV container is analyzed to optimize the performance of the cluster for computer vision applications while accommodating numerous IoT sensors. Through various experiments, this doctoral dissertation extensively analyzes the performance of various containerized AI vision applications at the edge and demonstrates how to boost their efficiency by carefully integrating container technologies into IoT and ARM-based edge computing environments.
In conclusion, this doctoral dissertation improves the understanding of integrating AI vision into IoT and edge computing, paving the way for deploying AI vision capabilities in real-time, edge-driven smart environments.Computer Scienc