48440 research outputs found
Sort by
The Gender Division in Grading and Groupwork
No full textThe high school classroom is often an important gateway for students that brings them in contact with subjects they might pursue in higher education. The issue, however, is that the traditional system of grading and lesson planning may fail to produce sufficient intrinsic motivation for students to pursue topics further, instead encouraging students to work for a grade and not for a love of learning. To change this system, however, it is important to factor in how students will respond to a variety of approaches, especially when considering the gender of the student. My work was to investigate the differences in responses across gender groups in a high school physics classroom to determine if a suitable alternative system for grading and classwork could be equitably achieved
Development of Open-source Educational Tools for Security Analysis of Integrated Circuits (SCAPEgoat)
No full textSide channel attacks exploit physical leakages, such as power consumption, to extract crypto- graphic secrets from embedded devices, bypassing mathematically secure encryption algorithms such as AES. This project enhances the accessibility and versatility of Side-Channel Analysis (SCA) research by extending the open source SCApegoat Python library to support the Pico- scope 2208B digital oscilloscope. By integrating the PicoSDK with SCApegoat, we enable power trace capture independent of specialized capture hardware such as the ChipWhisperer, though our demonstration uses the board for communication purposes. Our work involved developing a modular Python interface for the Picoscope, which simplifies signal acquisition, synchronization, and data storage. We demonstrated the effectiveness of our setup by successfully conducting cor- relation power analysis attacks on an AES-128 implementation. By lowering the cost and technical barriers to entry, this project expands the opportunities for researchers and students to explore hardware vulnerabilities and develop more secure embedded systems
Trading the Currencies (Forex) Market with Proprietary Firm Capital
No full textThis 21-week forex trading project explores how participants can manage and diversify a 100,000 accounts from FTMO and the 5ers, achieving 10% and 5% specific profit targets across initial and verification phases for the simulated $200,000 total capital allocation. The final funded phase realized significant gains, culminating in a profit management fee payout request
Designing of a Disaster Relief Drone
No full textWith robotics and artificial intelligence advancing over time, industries are becoming more automated. One aspect that is becoming standard in the industry are unmanned aircraft systems that can deliver supplies to difficult to reach areas. The packages could be used for general purpose or designed to deliver specific products, such as medical consumables, vaccines, blood samples and prescription drugs. This Major Qualifying Project reviews aspects of package delivery drones and adjusts them to deliver medical supplies to disaster relief areas. There are several things that need to be analyzed and can be implemented into the drone design to deliver medical supplies, such as the shape and placement of motors used on the drone, mechanisms for grippers to secure onto the package, and how to protect the internal supplies in the package. It is also important for what is in the package and the possibility that they are materials that are easy to contaminate, so making a sterile environment for these products is important. This project results in a prototype quadcopter drone that is 13 pounds and can reach an altitude of up to 2500 ft., a bistable mechanism involved with carrying and securing a package, as well as a 5-to-8-pound payload able to transport sensitive medical products. The design and manufacturing of the prototype are based on state, federal, and international standards. The project discusses engineering working drawings, the simulations of impact resistance on the drone, and testing of the prototype. The conclusions that this project reaches give several recommendations on important design aspects to be considered when creating aerial package delivery drones, ways to be cost effective during the manufacturing process, as well as potential ways for this drone to be used in the future
Modeling, Transfer, and Reasoning in Biomedical AI: From Clinical Trial Prediction to Interpretable Question Answering
No full textAdvances in artificial intelligence (AI) have opened new opportunities for understanding complex relationships among drugs, diseases, and treatment protocols, as well as for enhancing knowledge discovery from biomedical literature. Recent breakthroughs such as AlphaFold for protein structure prediction, BioGPT and Galactica for biomedical language understanding, and Med-PaLM for clinical reasoning have demonstrated the transformative potential of AI in biomedicine. Yet, despite these successes, most existing systems remain domain-specific, heavily data-dependent, and limited in their ability to integrate heterogeneous modalities or provide interpretable, evidence-grounded insights. However, key challenges remain in (1) integrating heterogeneous biomedical modalities for accurate clinical trial outcome prediction, (2) optimizing large-scale models to efficiently learn cross-modal representations, and (3) enabling interpretable reasoning to support transparent biomedical question answering. This dissertation addresses these challenges through three complementary research directions that together advance the modeling, optimization, and interpretability of biomedical AI systems. First, we propose Mode Experts Cross-Attention for Clinical Trial Outcome Prediction (MEXA-CTP), a novel model architecture that employs modality-specific experts and cross-attention mechanisms to capture fine-grained interactions among drug, disease, and protocol information. Optimized via Cauchy and contrastive losses, MEXA-CTP effectively models drug–disease–protocol dependencies without hand-crafted fusion structures. To further enhance scalability and transferability, we develop CLaDMoP (Learning Transferrable Models from Successful Clinical Trials via LLMs), which leverages large language models (LLMs) to encode complex clinical trial text and link it to a lightweight drug–molecule branch through a novel grouping-block-based multi-level fusion technique. CLaDMoP enables efficient cross-modal representation learning and knowledge transfer across diverse clinical tasks. To facilitate this research, we construct the Successful Clinical Trial (SCT) dataset—a domain-specific benchmark that supports pre-training and evaluation for clinical trial outcome prediction. Finally, we extend our investigation toward biomedical question answering, proposing PubMed Reasoner (Dynamic Reasoning-based Retrieval for Evidence-Grounded Biomedical Question Answering), a training-free, multi-stage agent framework inspired by the reasoning process of human researchers. PubMed Reasoner performs iterative query refinement, reflective retrieval, and evidence-grounded response generation, improving factual grounding and interpretability in biomedical reasoning. Collectively, this dissertation establishes a unified research trajectory toward trustworthy, efficient, and interpretable biomedical AI. By bridging small-model design for targeted multimodal integration, large-model transferability for scalable clinical understanding, and reasoning-based interpretability for transparent knowledge discovery, this work contributes a holistic framework that advances the development of intelligent, explainable, and generalizable systems for biomedical research and healthcare applications
GaitNet: Greedy, Acyclic Quadruped Gait Generation
No full textQuadruped locomotion across unstructured terrain remains a longstanding challenge in robotics, requiring controllers that can adapt dynamically to complex and unpredictable environments. Traditional approaches rely on model predictive control or pre-defined gait schedules, offering stability but limited adaptability. Conversely, end-to-end learning methods enable agile and versatile behaviors but often sacrifice interpretability, safety, and real time feasibility. This thesis introduces GaitNet, a hybrid control framework that integrates a greedy, neural network-based gait planner with a traditional model-based controller to generate dynamic, acyclic locomotion for quadruped robots. The proposed system consists of two primary components: a Footstep Evaluation Network that learns terrain-aware footstep cost maps, and GaitNet, a reinforcement learning-based gait selector that ranks and executes feasible footstep actions in real time. Together, these modules enable a balance between the adaptability of learning-based control and the robustness of analytical motion planning. Trained in NVIDIA Isaac Lab using parallel GPU simulation, the system is evaluated across a range of terrain difficulties and commanded velocities. Experimental results show that GaitNet achieves a 77.7% mean survival rate in challenging terrain, outperforming a single-leg motion planner baseline by more than a factor of two. Ablation studies further reveal that dynamic swing duration offers limited benefit in static environments, while removing pre-computed footstep costs leads to improved performance and more efficient gait patterns. These findings highlight the advantages of direct policy learning for coordinated, multi-leg motion without over-reliance on heuristic priors. Overall, this work demonstrates that greedy, neural network-based planning can produce dynamic, non-gaited quadruped motion. The presented framework bridges the gap between reactive learning and structured control, providing a foundation for future research toward fully autonomous, terrain-adaptive legged locomotion in real-world settings
Leveraging Clippy on an Embedded Rust Codebase
No full textSoftware delivered to the United States Department of Defense (DoD) has to abide by strict requirements in order for it to be utilized in mission-critical environments. In this work, I use a popular static analysis technique, known as lint, to cover requirements provided by the DoD for Magnetite, a Rust based operating system. To do this, I mapped relevant lints from Rust’s built-in linting tool, Clippy, and the Rust compiler to as many requirements possible. For the remaining requirements that could not be covered through existing lints, I developed my own Clippy toolchain with custom lints to satisfy requirement needs. This custom toolchain serves as a starting point for the Magnetite team in covering DoD requirements in the future
Assessing Community Engagement Strategies in Cambridge
No full textCombined Sewer Overflows (CSOs) occur when rainwater overwhelms sewer systems that carry both stormwater and sewage, releasing untreated sewage into local waterbodies and harming health and the environment. In Cambridge, MA, many residents are unaware of CSOs’ impacts and proposed solutions. We worked with the City of Cambridge to create an outreach strategy using interviews, visual elicitation, website analysis, and a focus group. We found that proper outreach has varied avenues of communication, in-person engagement, multilingual and accessible materials, and transparent feedback loops. We recommend improving website design, using clear graphics, leveraging community networks, and being open about community feedback to build engagement
Enhancing Community Engagement and Altruistic Education for Children in Taiwan
No full textThe goal of this project was to assist the Well Planet Foundation in Taiwan to improve its outreach and engagement efforts for ChildGood LAB, an altruistic education program targeting affluent families. To realize this goal, we conducted semi-structured interviews, surveys, and archival research to understand what motivates or prevents participation. We found that time constraints and a lack of academic relevance often discourage enrollment. As a result, we proposed an outreach strategy and an approach that frames the program as both a tool for personal growth and a valuable addition to students’ academic and extracurricular profiles. These findings aim to help Well Planet better align with parental priorities and increase long-term civic engagement
Evewave Zero Waste Startup
No full textEVEWAVE, a Copenhagen-based startup, aims to reduce the time, complexity, and cost in the supply chain with EVE1–a reusable pallet integrated with IoT to reduce waste and optimize logistics. This study evaluated the current packaging industry, including changing EU regulations, plastic waste generation, and lack of data. Through interviews, field observations, hands-on testing, and target audience analysis, we identified key areas for improvement. Using these methods, we created three deliverables: product refinement suggestions, a marketing brochure, and a preliminary website plan. Our findings support EVE1 as technology that helps build an efficient and sustainable supply chain system. They also provide insights about user experience and convey EVE1’s value to stakeholders