48440 research outputs found
Sort by
Rethinking RNN Training: A Dynamical Systems Approach to Training, Optimization, and Distillation of LLMs
No full textThere has been a growing consensus among the Computer Science community regarding the failings of RNNs. These machine learning model architectures are often seen as being difficult to work with for most real-world problems, with issues like vanishing and exploding gradients problems making it difficult for them to converge using normal training methods. The goal of this research is to study methods used in the field of Dynamical Systems Theory and apply those methods to create a generalized dynamical system architecture that can close the performance gap between RNNs and other neural networks as well as provide alternative trade-offs not seen in other neural network models. In particular, we propose to compare our techniques to transformed based LLM models such as Falcon by using distillation to replicate the performance seen in modern Transformers
Robust Depth Estimation via Monocular and Focal Stack Data Fusion for Mobile AR
No full textMetric depth estimation plays an important role in mobile augmented reality (AR). With accurate metric depth, we can achieve more realistic user interactions, such as object placement and occlusion-aware rendering. While specialized hardware like LiDAR demonstrates its promise, its restricted availability—only on selected high-end mobile devices and performance limitations, such as detection range and environmental sensitivity, make it less ideal. Monocular depth estimation, on the other hand, relies solely on mobile cameras, which are ubiquitous, making it a promising solution for mobile AR. In this thesis, we investigate the challenges of achieving accurate metric depth estimation in mobile AR and propose a novel solution called HYBRIDDEPTH. Through an evaluation of four state-of-the-art monocular depth estimation models on a recently introduced dataset (ARKitScenes), we identified three types of challenges: hardware, data, and model-related challenges. To address scale ambiguity, generalization, and hardware heterogeneity challenges, we propose HYBRIDDEPTH, a robust depth estimation pipeline that addresses key challenges in depth estimation, including scale ambiguity, hardware heterogeneity, and generalizability. HYBRIDDEPTH leverages the focal stack, data conveniently accessible on common mobile devices, to produce accurate metric depth maps. By incorporating depth priors afforded by recent advances in single-image depth estimation, our model achieves a higher level of structural detail compared to existing methods. We test our pipeline as an end-to-end system with our developed mobile client to capture focal stacks, which are then sent to a GPU-powered server for depth estimation. Comprehensive quantitative and qualitative analyses demonstrate that HYBRIDDEPTH outperforms state-of-the-art (SOTA) models on common datasets such as DDFF12 and NYU Depth V2. HYBRIDDEPTH also shows strong zero-shot generalization. When trained on NYU Depth V2, HYBRIDDEPTH surpasses SOTA models in zero-shot performance on ARKitScenes and delivers more structurally accurate depth maps on Mobile Depth. For example, HYBRIDDEPTH achieves a 10.5% and 6.1% improvement of MSE and RMSE on DDFF12, and 6.5% and 7.1% improvement of RMSE and AbsRel on the NYU Depth V2 dataset, compared to DFV. Moreover, our zero-shot performance on the ARKitScenes dataset shows a 43% improvement compared to SOTA methods
Empowering Educators, Impacting Students: Moving Research to Practice
No full textLearning science research has made impactful strides in understanding the cognitive processes underlying learning and instruction. In the dynamic landscape of education, integrating learning science research into pedagogical instructional design is a crucial undertaking. This dissertation investigates the challenges and opportunities associated with translating learning science research into education practice. Grounded in the theoretical work of learning sciences, this dissertation examines the implementation of embodied cognition, desirable difficulties, and productive struggle. It highlights the complex interplay between research evidence and practitioner expertise. This study explores promising pathways to bridging the gap between research and practice, including professional development opportunities, collaborative networks, and instructional change. This project contributes to a deeper understanding of how learning science research can inform and improve educational practices. By addressing the obstacles and benefits of adopting evidence-based strategies, educators can enhance student learning and foster a culture of continuous improvement in educational settings
Generation AI Project-Based Learning (PBL) Workshop
No full textFAST developed the Generation AI program to help students enter the AI field. To improve their program, they wanted a project-based learning (PBL) workshop to bridge the gap between theory and practice. After researching and conducting interviews, we identified key aspects of effective PBL workshops, which we utilized to develop a four-module Practical Applications of Machine Learning workshop. Supporting materials like presentations, assessment tools, and rubrics were provided for FAST to replicate for their students and other Armenian high schools
PMKS+: An Application for Generating and Analysing Planar Linkages
No full textThe PMKS+ (Planar Mechanism Kinematic Simulator Plus) MQP focuses on enhancing a web application for simulating planar linkages with a single degree of freedom. The aim is to restore legacy features while adding new functionality and improving usability. Accomplishments include implementing three-position synthesis, six-bar linkage creation, and a new method for specifying coupler positions. Other additions were: optimized animation; integrated animation bar; URL generation; revised kinematic calculations; and length and angle display. The user interface was redesigned with a refined panel layout and smoother cross-panel behavior. These updates recover previous features and make PMKS+ more intuitive, transforming it into a more effective tool for kinematic analysis and linkage design
EduVentures: Developing video and educational materials for the Natural History Museum
No full textWorking with EduVentures Trust in partnership with the National Museum of Namibia (NMN), our goal was to create educational content about natural history museums that will be shared with the public. We used a combination of interviews and surveys to support the creation of our educational materials. Curators, teachers, students, tourists, and members of the public revealed factors that contribute to the lack of knowledge surrounding natural history, such as working time, social class, grade level, gender, and more. Our final products include an engaging video documenting the process of collecting species for the museum, an informative presentation highlighting each department and educating about the difference between natural and social history, and updated brochures for each department that are educational and attractive. We recommend further research that delves deeper into the factors we discovered to further ensure that equitable and quality education can be provided for all
Investigation of Large Language Models in Generating Accurate Identification of Molecular Machines Across Model Organisms
No full textMolecular machines, protein complexes in particular, play a key role in the everyday functions of all living organisms. Despite the significance of molecular machines, information about their composition remains scattered across articles and databases, and each database is reliant on experimental methods for validation for their data. However, with the ever-growing popularity of generative AI, Large Language Models have been proposed as a potential method in order to compile information about protein complexes into a singular dataset for widespread scientific use. By using models from OpenAI, Anthropic and Perplexity, implementing prompt engineering techniques, and optimizing the workflow, we were able to formulate a pipeline to generate protein composition of the queried proteins. It was found that the few-shot and contextual prompting techniques were more accurate with a partial accuracy of 65%, and the most proficient model was the GPT-4o. Our results show potential that LLMs can be valuable assistants in creating databases for widespread scientific use. As this was an exploratory study, future work can build upon ours by expanding the sample size, testing prompt hallucinations, and implementing protein-protein interactions into the database
Game On: An Interactive Sports System For Acquired Brain Injury Rehabilitation
No full textIntegrating sports-related exercise in the physiotherapy process creates a unique opportunity for therapists and individuals with acquired brain injury. The Interactive Sports for Rehabilitation System was designed with multiple subsystems, to enhance the hand-eye coordination, strength, cognitive function, and precision of upper limb movements of individuals with acquired brain injuries. Targets, lights, sensors, and software that went into the construction of the system all work together to create a dynamic rehabilitation process that focuses on improving physical and cognitive abilities
Sustainable Energy for Smarter Irrigation: Developing a Portable Solar-Powered Irrigation System for Tanzania and the Developing World
No full textThis project aimed to address agricultural water needs in Tanzania by developing a solar-powered irrigation system based on three key design principles: portability, affordability, and fluidity. Guided by survey and interview data, we created a pump that meets farmers’ needs while upholding these principles. The final design reduces reliance on non-renewable energy and supports sustainable agriculture. This project has the potential to benefit farmers in Tanzania and other developing regions facing similar challenges
Sniffing CANBUS Signals using Inductive Sensing and Software Defined Radios
No full textThis project aims to bring to light the possibility of using inductive sensing to sniff messages from a vehicle’s Controller Area Network (CAN) without ever having a physical connection to the system. This is achieved by firstly creating a custom vehicle test-bench to simulate the CAN system and then using an inductive probe to measure the electromagnetic fields emitted by the communication wires. With the use of a Software-Defined Radio (SDR) technology and Python signal processing libraries, this report evaluates the effectiveness of this new wireless inductive sensor attack and demonstrates that signals traveling through the CAN network can be eavesdropped on by unauthorized third parties