Portfolio

Natural Language Processing and Deep Learning

Epidemiology

Sports Analytics

Analysis and Fine-Tuning of an SNLI Trained Model using Static Dataset Cartography and Out of Domain Datasets

Abstract: An ELECTRA-small Natural Language Inference (NLI) model was trained based on the Stanford NLI dataset. Two methods are proposed for model improvement: 1) static dataset cartography and 2) training on Out of Domain (OOD) datasets. It is shown that training on ambiguous datasets offers the greatest advantage in fine tuning of the baseline model. Furthermore, training on small number of OOD examples can cause drastic increase in model performance on overall OOD datasets.

A State-Based Agent for Automated Goal Scoring in SuperTuxKart Ice Hockey

Abstract: This paper presents an agent that was trained using imitation learning on video game state data with the objective to automate goal scoring in SuperTuxKart Ice Hockey. A four-layer deep neural network was used to train the model on training data obtained from the teacher agent. It is shown that the trained agent closely matches the agent that it was trained to imitate (teacher), scoring 84% as many goals as the teacher agent..

COVID-19 Actual Prevalence in America Adjusted for known COVID-19 Mortality and for Excess Mortality during the Pandemic Period

Abstract: Regional and temporal variation in COVID-19 test availability have limited real estimates of COVID-19 prevalence over time. Better estimates could orient EDs to plan for surges of patients. Reports of seroprevalence of immune response to COVID-19 in American communities imply between 6 and 24 fold higher rates of infection than the official RT-PCR based case counts. We used the known deaths attributed to COVID-19 and the excess deaths [4] in America during the pandemic and fit the data to reported local timed cumulative case load (immunoprevalence) to derive upper and lower estimates of the true case fatality rate for COVID-19 in America.

Friends, Family and Sustainability: Lessons from a Novel Model for a New COVID-19 Pandemic

Abstract: SARS-CoV-2 has led to urgent and renewed interest in epidemiological modeling of novel infectious pathogens and interventions to curb them while awaiting vaccines and other effective treatments. We propose and describe a novel stochastic approach to SIR with subpopulations and ability to derive estimates of our uncertainty, both with respect to the model itself and the variables dependent on unknown characteristics of the pathogen modeled. Our model adds new value by accounting for subpopulations within a main population and for differential interactions within and between these subgroups, leading to different opportunities for infection between subgroups. We study various potential interventions during a pandemic as represented by their impact on transmission in subgroups in our model within simulations. We measure total cases over time, proportion of a populace requiring intensive care over time, and peak ICU utilization. We show convergence of our model with known models mathematically and validate our model with historical American influenza data. We then model various interventions to discern which are most effective. We find that: 1) Implementing excessively restrictive social distancing measures can have diminishing and even negative returns, particularly if these are only temporary and there are not ongoing modulated effects in effective . 2) Reducing the spread of a novel infectious disease within households and among family and friends is the most effective means by which to reduce the total infection rate. 3) To reduce the total number of infections, effective must be lowered in a sustainable fashion..

Performance Based Analysis for Unbiased Ranking of College Football Teams

Abstract: Coming soon