Internet use has been identified by researchers as having the potential to rise to the level of addiction, with associated increased rates of anxiety and depression. Identifying cases of problematic internet usage currently requires evaluation by an expert, however, which is a significant impediment to screening children and adolescents across society. One potential solution is to rely on data that is more easily and uniformly collected: the kind collected by a family physician, a simple survey, or by a smartwatch. The research question this project sets out to answer is: “Can we predict the level of problematic internet usage exhibited by children and adolescents, based on their physical activity and survey responses?”

This project was designed to investigate the potential relationship between AP exam performance and the presence of nearby universities. It was initially hypothesized that local (especially R1/R2 or public) universities would contribute to better pass rates for AP exams in their vicinities as a result of their various outreach, dual-enrollment, tutoring, and similar programs for high schoolers. We produce a predictive model that uses a few features related to university presence, personal income, and population to predict AP exam performance.

Modeling the relationship between biogeochemical layers and chlorophyll density
The distribution and density of chlorophyll in the ocean are critical indicators of marine primary productivity, which influences the global carbon cycle, marine food webs, and climate regulation. Biogeochemical and physical ocean properties, including nutrient availability, light penetration, water temperature, salinity, and ocean currents influence chlorophyll density. Understanding and accurately modeling these relationships is essential for predicting the impacts of environmental changes on marine ecosystems and for managing oceanic resources effectively. We plan to combine multiple Copernicus Marine Datasets to model the chlorophyll density based on the biochemical and physical properties of the ocean.

RivusVox Editor: the world's first near-live zero-shot adaptive speech editing system

The ReCANVo dataset consists of ~7k audio recordings of vocalizations from 8 minimally-verbal individuals (mostly people with developmental disabilities). The recordings were made in a real-world setting, and were categorized on the spot by the speaker's caregiver based on context, non-verbal cues, and familiarity with the speaker. There are several pre-defined categories such as selftalk, frustrated, delighted, request, etc., and caregivers could also specify custom categories. Our goal was to train a model, per individual, that accurately predicts labels and improves upon previous work.

We train several different combinations of models of the form “Feature Extractor + Classifier”. For extracting features from audio data, we use two deep models (HuBERT and AST) each with pre-trained weights, as well as mel spectrograms. As classifiers, we use a 4 layer CNN-based neural network (for mel spectrograms), NNs with fully-connected layers (for features coming from deep models), and more.

Throughout history, composers and musicians have borrowed musical elements like chord progressions, rhythms, lyrics, and melodies from each other. Our motivation for this project is born of a fascination with this phenomenon, which of course extends to less legal examples like unconsciously or intentionally copying the work of another. Even famed and highly regarded composers like Bach, Vivaldi, Mozart, and Haydn are not innocent of borrowing from their contemporaries or even recycling their own works. Similarly, in 2015, in a high-profile court case, defendants and artists Robin Thicke and Pharrell Williams were ordered to pay millions of dollars in damages for copyright infringement to Marvin Gaye's estate, considering they borrowed from Gaye’s "Got to Give it Up" when writing their hit "Blurred Lines." Our project aimed to use deep learning to assess the similarity between musical clips to potentially establish a more robust and empirical way to detect music plagiarism.

Knowing where to go to find customers is the most important question for taxi drivers and ride hailing networks. If demand for taxis can be reliably predicted in real-time, taxi companies can dispatch drivers in a timely manner and drivers can optimize their route decision to maximize their earnings in a given day. Consequently, customers will likely receive more reliable service with shorter wait time. This project aims to use rich trip-level data from the NYC Taxi and Limousine Commission to construct time-series taxi rides data for 63 taxi zones in Manhattan and forecast demand for rides. We will explore deep learning models for time series, including Multilayer Perceptrons, LSTM, Temporal Graph-based Neural Networks, and compare them with a baseline statistical model ARIMAX.

arXiv is the largest open database available containing nearly 2.4 million research papers, spanning 8 major domains covering everything there is to understand from the tiniest of atoms to the entire cosmos. A large language model (LLM) having access to such a dataset will make it unprecedented in generating updated, relevant, and, more importantly, precise information with citable sources.

This is exactly what we have done in this project. We have refined the capabilities of Google’s Gemini 1.5 pro LLM by building a customized Retrieval-Augmented Generation (RAG) pipeline that has access to the entire arXiv database. We then deployed the entire package into an app that mimics a chatbot to make the experience user-friendly.

The idea of the project is to predict high glucose spikes from continuous glucose data, smartwatch data, food logs, and glycemic index. The dataset consists of the following:
1) Tri-axial accelerometer data (movement in subject)
2) Blood volume pulse
3) Intestinal glucose concentration
4) Electrodermal activity
5) Heart rate
6) IBI (interbeat interval)
7) Skin temperature
8) Food log
Data is public in: https://physionet.org/content/big-ideas-glycemic-wearable/1.1.2/#files-panel

The nocturnal soundscapes of late summer and autumn are replete with the familiar chirps, trills, and buzzes of singing insects. But these cryptic performers often remain anonymous and underappreciated.

The goal of this project was to build machine learning models to identify the presence of insects in sound files and to coarsely categorize the sounds as crickets, katydids, or cicadas.

Both Support Vector Classifiers and Convolutional Neural Networks were able to identify insects songs to the broad categories of cricket, katydid, and cicada with 90% accuracy or higher.

In the future, similar, more sophisticated models could be applied to filtering large volumes of passively recorded audio from ecological studies of insects and could power apps that identify insect songs to the species level.

This project involves the investigation and evaluation of different methodologies for retrieval for use in RAG (Retrieval-Augmented Generation) systems. In particular, this project investigates retrieval quality for information downloaded from employee subreddits. We investigated the impacts of using clustering, multi-vector indexing, and multi-querying in advanced retrieval methodologies against baseline naive retrieval.

Groundwater is a critical source of water human survival. A significant percentage of both drinking and crop irrigation water is drawn from groundwater sources through wells. In the US, overuse of groundwater could have major implications for the future and forecasting groundwater can be useful in understanding its impact. Building on historical data for four wells, together with surface water and weather data, in Spokane, WA, we construct and evaluate machine learning models that forecast groundwater levels in the area.

Most existing music recommendation systems rely on listeners to provide seed tracks, and then utilize a variety of different approaches to recommend additional tracks in either a playlist-like listening session or as sequential track recommendations based on user feedback.

We built a playlist recommendation engine that takes a different approach, allowing listeners to generate a novel playlist based on a semantic string, such as the title of desired playlist, specific mood (happy, relaxed), atmosphere (tropical vibe), or function (party music, focus). Using a publicly available dataset of existing playlists (https://www.aicrowd.com/challenges/spotify-million-playlist-dataset-challenge), we combine a semantic similarity vector model with a matrix factorization model to allow users to quickly and easily generate playlists to fit any occasion.

Preterm birth is a primary cause of infant mortality and morbidity in the United States, affecting approximately 1 in 10 births. The rates are notably higher among Black women (14.6%), compared to White (9.4%) and Hispanic women (10.1%). Despite its prevalence, predicting preterm birth remains challenging due to its multifaceted etiology rooted in environmental, biological, genetic, and behavioral interactions. Our project harnesses machine learning techniques to predict preterm birth using electronic health records. This data intersects with social determinants of health, reflecting some of the interactions contributing to preterm birth. Recognizing that under-representation in healthcare research perpetuates racial and ethnic health disparities, we take care to use diverse data to ensure equitable model performance across underrepresented populations.

Alzheimer’s disease (AD) is one of the most common types of dementia and frequently affects the elderly. Electroencephalography (EEG) is a non-invasive technique to measure the brain activity using external electrodes and may help provide improved diagnosis of AD. In this project we use power spectrum of EEG to build a robust machine learning classifier which predicts whether a patient has Alzheimer's or is healthy. We vastly improve upon existing models in the literature by using modified features compared to the ones used in literature.

Fingertip pulse oximeters are the current standard for estimating blood oxygen saturation without a blood draw, both at home and in healthcare settings. However, pulse oximeters overestimate oxygen saturation, often resulting in ‘hidden hypoxemia’: a patient has hypoxemia (dangerously low oxygen saturation), but the oximeter returns a healthy oxygen value. Unfortunately, oximeter overestimation of oxygen saturation is exacerbated for patients with darker skin tones due to light-based oximeter technology. This results in Black patients experiencing hidden hypoxemia at twice the rate of white patients. By combining pulse oximeter readings (SpO2) with additional patient data, we develop improved methods for estimating arterial blood oxygen saturation (SaO2) and identifying Hidden Hypoxemia. The predictions of our models are more accurate than pulse-oximeter readings alone, and remove the systematic racial inequity inherent in the current medical practice of using oximeter readings alone.

Our problem is to accurately predict maize field centers in Africa using very low resolution satellite images. The dataset contains many disparate entries including two satellite imagery, one with higher spatial res (Planet) and one with higher temporal resolution and wavelength coverage (Sentinel-2), partial metadata about the crop fields including estimated yield, size, and subjective quality of the measurement. We are employing CNN based image segmentation models to compute displacement vector from the image center.

Acronym/Abbreviation (short form) disambiguation is one of main challenges when using NLP methods to uderstand medical records. While this topic has long been studied, it is still a work in progress. Current strategies often involve having manually curated datasets of abbreviations and train classifiers. The main problem of that approach is that curated datasets are sparse and don't include all the short forms. In Dec 2020, a paper came out where they created a large dataset of short forms as one of their steps in their pipeline to pre-train models. The goal of our project would be to build upon their short form disambiguation piece and create a tool to disambiguate a medical short form using its context. Example of the usage of our tool: original_sentence = "The patient states that she has had dizziness, nausea, some heartburn, and some change in her vision. She is gravida 6, para 4, AB 2. She has no history of adverse reaction to anesthesia." AB could stand for "abortion", "ankle-brachial", "blood group in ABO system", "A, B lines in Kerley lines". disambiguated_sentence = "The patient states that she has had dizziness, nausea, some heartburn, and some change in her vision. She is gravida 6, para 4, abortion 2. She has no history of adverse reaction to anesthesia."

We built a podcast recommendation engine that suggests episodes to a listener based on either a previous episode that they've heard or an episode description that they can input with freeform text entry.

We built a movie finder app that allows a user to enter some details they remember about a movie (along with some optional filter info on the genre and release year) and then predicts what movie the user is thinking of. To solve this NLP problem, our tool uses an embed-and-rerank model. We have precomputed vectorizations of movie plot information for the approximately 34,000 movies in our dataset.

Our model’s first step is to vectorize the user’s query and do a fast comparison to find the 100 closest plot vectors. Then it reranks these top 100 closest plots, performing a more thorough comparison using a neural network that semantically compares the plot fragments with the original query. Finally, we output the 10 movies which show up at the top of this new ranking.

Timely identification of safety-critical events, such as gunshots, is of great importance to public safety stakeholders. However, existing systems only deliver limited value by not classifying additional urban sounds. We perform classification of environmental sounds to detect safety-critical events, in particular gunshots, and provide information on first-response via siren detection. We also engineer general features for off-line classification tasks and demonstrate how this system can provide value to additional stakeholders in the film and television industry.

The NHL Stanley Cup is the oldest existing trophy to be awarded to a professional sports franchise in North America, and often considered “the hardest trophy to win in professional sport.” Using just regular season data, we want to know, can we predict who is going to win the Stanley Cup?
We collected data from each team, as well as data from every player in over 20,000 games going back to 2005. Using this data, we made an ensemble model using logistic regression, AdaBoost, random forests, and a neural network, which were able to predict playoff data with up to 70% accuracy - above the theoretical threshold reported in the literature of 62%.