7903 research outputs found
Sort by
Development of Sustainable Nano-Engineered Cementitious Composites Reinforced with Graphene Oxide from Invasive Seaweed
Full text linkThis research explores the integration of graphene oxide (GO) derived from invasive sargassum seaweed into cementitious composites to enhance concrete performance while addressing critical environmental and sustainability challenges. With the dual goal of reducing the carbon footprint of cement and repurposing harmful sargassum blooms in Puerto Rico, the research investigates the viability of sargassum-derived GO (SGO) as a partial cement substitute in mortar mixes. The study is motivated by the urgent need to lower cement-related CO₂ emissions—currently contributing nearly 8% of global totals—and by the ecological and economic burdens caused by the accumulation of sargassum along coastlines. Graphene oxide, with its exceptional mechanical and microstructural properties, offers the potential to improve concrete’s strength, durability, and workability. This research adopts a lab-based methodology involving the controlled batching of mortar samples with varying concentrations of both industrial and sargassum-derived GO. Key experimental procedures include compressive strength testing using a universal testing machine, flow table tests for workability, thermogravimetric analysis (TGA) for thermal stability, and X-ray diffraction (XRD) to examine mineralogical changes. Findings indicate significant improvements in both early-age and long-term compressive strength of GO-enhanced mixes compared to controls. The optimal replacement dosage was identified as 0.05% GO by weight of cement, consistent with prior literature and confirmed through experimental results. While the conversion of sargassum to GO is not directly within the study’s scope, the project builds on established pyrolysis methods and emphasizes the environmental advantages of using rapidly renewable biomass. The thesis focuses on optimizing mix designs, understanding material behavior, and comparing the performance of SGO-enhanced concrete to that of conventional GO and control mixes. Ultimately, this research seeks to contribute to the development of greener construction materials while promoting a circular economy and offering potential socio-economic benefits to sargassum-affected communities
Tiny-TLIC: Lightweight Transformer Based Lossy Image Compression
Full text linkThis thesis presents a comprehensive study on image compression, beginning with a review of both traditional and learning-based methods. Building upon the conventional Variational Autoencoder (VAE) framework, we explore the integration of Vision Transformer (ViT) architectures into the image compression pipeline to leverage their global modeling capabilities. Based on this, we propose Tiny-TLIC (Tiny Transformer-Based Lossy Image Compression), a lightweight yet effective model that introduces two novel components: Integrated Convolution and Self-Attention (ICSA) for enhanced feature representation, and a Multistage Context Model (MCM) for improved entropy estimation. To validate our approach, we train both a baseline Transformer-Based Image Compression (TIC) model and the proposed Tiny-TLIC. Extensive evaluations on the Kodak and other benchmark datasets demonstrate that Tiny-TLIC achieves competitive performance with significantly reduced model size, highlighting its potential for practical applications in resource-constrained environments
Adaptive Lambda Scheduling in the DREAM Super-Resolution Model
No full textDiffusion-based super-resolution (SR) models often suffer from a training–inference discrepancy: during training, they denoise images conditioned on ground-truth high-resolution targets, but during inference, they rely solely on self-generated predictions. The DREAM framework (Diffusion Rectification and Estimation-Adaptive Models) addresses this mismatch by blending model predictions with ground-truth supervision through a time-dependent weight, λt .
In this work, we investigate how the scheduling of λt affects SR quality. Specifically, we evaluate three λt scheduling strategies: a power-law schedule (used in the original DREAM), a cosine-based curve, and a sigmoid-shaped curve. Each defines a distinct way to balance reliance on ground truth versus the model’s own predictions throughout the diffusion process.
We conduct 16× super-resolution experiments on the CelebA-HQ dataset, measuring fidelity with PSNR and SSIM, and perceptual quality with LPIPS and FID. Results demonstrate that λt scheduling significantly influences the trade-off between pixel-level accuracy and perceptual realism. Among the tested strategies, cosine scheduling consistently provides the best balance. These findings highlight the critical role of adaptive λt design in improving the performance of diffusion-based SR systems
Accelerating LLM Inference with Smart NIC Tokenization and Caching
No full textTokenization is a critical preprocessing step in machine learning pipelines, particularly for generative AI models such as large language models (LLMs). Despite significant advances in training and inference technologies, tokenization continues to be a notable performance bottleneck due to its CPU-bound nature and limited acceleration support. This thesis proposes NetTokenizer, a system designed to offload tokenization from CPUs to network data planes. Utilizing software frameworks such as the Data Plane Development Kit (DPDK) and hardware-based SmartNIC solutions, NetTokenizer substantially improves inference efficiency. Experimental evaluations demonstrate up to a 95.9% reduction in tail latency and a 5.73x throughput increase compared to conventional CPU-based methods. This research provides a comprehensive analysis and practical foundation for deploying in-network tokenization solutions to enhance ML inference pipelines
Empowering Tanzanian Education: Personalized and Accessible Test Preparation
No full textOnly 20% of students performed well enough on their secondary exams to continue their A-level studies. This challenge exists due to a lack of resources, classroom overcrowding, and absenteeism of the instructors. The project focuses on improving the pass rate for these national exams; we have built an intelligent quiz generating platform that helps Tanzanian students prepare more effectively for exams by meeting user needs including active recall, answer explanations, increasing difficulty, and filtered studying. Students can select the subject, form level, topic, and the difficulty level, and the platform provides different types of questions, including true/false, multiple choice, and free response, accordingly, along with more difficult incorrect answers. Furthermore, our platforms leverages AI to provide answer explanations to students, ensuring that students are not confused on why a specific answer is correct. Over the first 30 days after deploying, we have had 64 active users, and 1,600 page accesses
Notewise: An Interactive AI Tool for Music Theory Education and Composition Analysis
No full textThis paper presents Notewise, a cutting-edge VST3 plugin designed to promote music theory learning and compositional analysis on digital audio workstations (DAWs). Notewise bridges the gap in music composition software by delivering contextual and intelligent feedback on musical compositions in real-time. A modular Musical Instrument Digital Interface (MIDI) analysis platform drives Notewise, o↵ering versatile tools to assist users by recognizing and highlighting errors in the harmonic and rhythmic structures of their compositions. The context-aware nature and flexibility of the system makes feedback accommodating to multiple genres and styles, with composers improving their work while retaining control over the creative process. Unlike current solutions, Notewise is incorporated into the production process as a seamless component, offering dynamic yet non-obtrusive feedback. In its current form, users can choose between a recurrent neural network (RNN) or a deterministic analyzer to examine their works, offering customization and precision. With its innovative method, Notewise democratizes music theory learning by avoiding social, economic, and political obstacles, making learning materials available to everyone around the world. This thesis will discuss Notewise’s design, its innovation in music theory learning, and its potential in the future to revolutionize music composition tools
Ergonomic Human Robot Handovers Using Surface Electromyography (sEMG) Sensors
No full textFor decades, robots have been kept in cages in industry. With the advances of collaborative robots and Artificial Intelligence (AI), there is a shift towards humans and robots working together. In this research, we propose an ergonomically friendly collaborative robotic cell that enables a human and a collaborative robot to work synergistically to assemble a mobile robot. The collaborative robot provides the parts while explaining the process through a computer, and the human co-worker follows the instructions to complete the assembly. The proposed collaborative robotic cell is evaluated in a user study to ensure that the handovers of the parts between the robot and the human happen ergonomically. Muscle activation data is gathered during human-subject experiments using surface electromyography (sEMG) sensors and post-experiment data through a survey. Three machine learning models, including 4-layer neural networks (4-layer NN), convolutional neural networks (CNN), and long short-term memory (LSTM), are developed to classify ergonomic and non-ergonomic status of muscle activation patterns using the Rapid Upper Limb Assessment (RULA) scores. The results show that the NN are capable of distinguishing the ergonomic status of the experiment with 80%, 83%, and 88% for LSTM, 4-layer NN, and CNN, respectively. The post-experiment questionnaire evaluating the human interaction with the collaborative robot showed positive experience, specifically on safety, usability, and trust in the system we developed. Our findings provide insights into the successful implementation and evaluation of human-robot collaboration in industrial assembly tasks. This research contributes to the growing trend of ergonomically centered robot design, focusing on reducing the risk of musculoskeletal disorders (MSDs) among human workers through collaborative robotics. Notably, the F1 scores of our models highlight the reliability of our ergonomic status classification. Overall, this study underscores the potential of cobots in creating safer and more efficient work environments while prioritizing human health and well-being
TCRAS: Traffic Control Risk Analysis System
No full textTraffic intersections represent critical points of conflict in urban transportation networks, with over 40,000 traffic-related fatalities occurring annually in the United States alone. Traditional intersection monitoring systems, based on timer controls and inductive loop detectors, lack the sophisticated detection capabilities needed to address modern traffic safety challenges. This thesis presents the Traffic Control Risk Analysis System (TCRAS), a low-cost, computer vision-based solution that democratizes access to advanced intersection monitoring capabilities. TCRAS leverages edge AI processing through Hailo neural network accelerators combined with open-source computer vision algorithms to provide comprehensive intersection analysis. The system performs real-time multi-class object detection and tracking of vehicles, pedestrians, and traffic lights, enabling sophisticated violation detection including red light violations, yellow light violations, jaywalking, congestion monitoring, and intersection blocking. The TCRAS Index provides a standardized 0-100 safety metric that synthesizes multiple violation types and efficiency factors into a single comparable score. The system was deployed to a real intersection for monitoring in Santa Clara, to prove its feasibility. This work demonstrates that sophisticated traffic analysis capabilities can be achieved without prohibitive infrastructure investments, potentially accelerating the adoption of vision-based safety systems in support of Vision Zero initiatives
3D Printing Filament Machine (REVERS3D)
No full textThe Reverse3D printer project aims to develop an innovative machine that recycles leftover 3D printed parts into new, usable spools, promoting sustainability in the 3D printing community. The project will create a new dimension of versatility when using 3D printing filament while still being user-friendly. The device will incorporate a grinding unit to grind up the parts, a heating element, an extrusion system, a cooling system, and a spooling system to create a continuous filament strand. Our ultimate goal is to grind the plastic to a size capable of being melted down and molded into a spool-like form suitable for being reused in a 3D printing machine
Bronco eVTOL: VFS 2024 - 2025 DVBF
No full textCommon vertical take-off vehicles, such as helicopters, have been useful for transportation, rescue, and fire-fighting applications. However, these vehicles contribute significantly to CO2 emissions and are extremely loud. Electric vertical take-off and landing vehicles ( e VTOLs) have been proposed as an alternative vehicle which can perform the same functions, without the costs associated with a helicopter. In this paper, we develop a conceptual design for an eVTOL which satisfies the requirements of the annual Vertical Flight Society\u27s Design-Vertical-Build-Fly Competition. We focus our conceptual design on three subsystems: Propulsion, Airfoils, and Controls. These subsystem designs are verified using Finite Element Analysis techniques, as well as Computational Fluid Dynamics simulations. Additionally, a preliminary control simulation using MATLAB /Simulink TM has been developed using these design choices. The next steps of the project will include detailed design of the tail, designing the payload dropping mechanism, and further developing the controls program. Long-term work includes manufacturing and testing of the aircraft in order to meet the competition deadline of April 1, 2025