Samuel Maley

Machine Learning Research Intern

• Investigated the effects of data augmentation on satellite imagery as a distinct data modality in object detection
• Utilized PyTorch, CUDA, Meta's Detectron2, and MMDetection to create data augmentation and model training pipelines for overhead image chips to compare each augmentation's effect on accuracy
• Trained several deep learning model architectures for object detection, including: Shifted Window (Swin) transformer, 101-layer Residual Network (ResNet101), and YOLO
• Performed dataset preprocessing to generate COCO style annotations for over 180,000 object instances across 2,800 images
• Automated satellite imagery chipping to create 256x256 pixel image chips from iSAID dataset

Data Science Research Assistant

• Researched various topics in video quesion-answering as part of the DARPA ECOLE project under Dr. Daisy Wang
• Contributed to DARPA's Environment-driven Conceptual Learning (ECOLE) by researching potential improvements to a few-shot classification pipeline for complex video actions
• Implemented Meta's Segment Anything Model (SAM) and Grouding-DINO to isolate and mask objects in videos from a single textual input
• Deployed state-of-the-art computer vision models to analyze object affordances in videos, enabling the prediction of potential action locations
• Created bash scripts utilizing FFMPEG to generate image and video test cases from the STAR dataset

Undergraduate Research Assistant

• Conducted comprehensive analysis of human-robot interactions to identify potential risks and design safeguards, collaborating closely with cross-functional teams to research emerging technologies
• Assisted in the design and use of EMG and IMU sensors to collect human participant data over 60+ hours of experiment data
• Utilized pandas and numpy to prepare collected EMG and movement data for statistical inference. Assisted in the design of a VR human-robot interaction data pipeline through the use of ROSBridge, Unity, and Nvidia Omniverse

Programming Tutor & Marketing Assistant

• Provided personalized computer science instruction to students aged 7-16
• Managed communications between parents
• Assist with marketing and collaborate with manager to engineer effective advertisement strategies
• Conducted cold calls and emails to expand business

University of Florida

Machine Learning on Tic-Tac-Toe Dataset

This repository contains the code for a machine learning project focused on building and evaluating various models on a tic-tac-toe dataset. The project involves implementing classifiers and regressors using libraries such as pandas, numpy, and scikit-learn, and demonstrates how these models can be used to predict optimal moves in a tic-tac-toe game.

Adversarial Attacks on Multimodal Entity Linking (MEL)

Multimodal Entity Linking (MEL) is a task aimed to match entities with the corrent types of data across different formats, such as text, images, and videos. The goal is to connect mentions of these entities to the correct entries in a structured knowledge base. However, data poisoning attacks, where attackers deliberately inject misleading or malicious data into the training set, can significantly compromise the performance of these systems. This project investigates the effects of these attacks on the state-of-the-art methods.

View Research Paper

MediGator: AI Healthcare Assistant

Developed a full-stack Flask/React healthcare application featuring a Firebase-integrated database, Clerk-powered user authentication, and an LLM-driven chatbot. Led a team of 4 as Scrum Master, implemented database methods, designed React components, and integrated Google Maps API for a provider locator tool.

YouTube Video Summarizer

This project utilizes OpenAI's ChatGPT API and a Youtube transcript API to generate summaries of an inputted YouTube video.

AI-in-Classics: Ancient Latin/Greek Temporal Sentiment Analysis Tool

Collaborated on a Flask/React application to analyze the sentiment of Latin and Greek words across ancient texts. Automated date estimation for 60,000 text segments using Wikipedia's API and LLM-powered scraping, while developing Flask methods and React components for seamless input and display.