Ryan Sawasaki

Hyperion SolarNet

A capstone project utilizing deep learning methods for automated detection of solar panel locations and their total surface area using aerial imagery. HyperionSolarNet consists of a two-branch model using an image classifier in tandem with a semantic segmentation model to provide an efficient and scalable method for detecting solar panels with reliable performance. This research project includes an application that provides users with an interactive mapping tool that visualizes the predicted solar panel mask overlay output from the classification and segmentation models.

• Neurips 2021
• Research Paper

• Website
• Web Application

• Presentation
• Github

Reading Comprehension and
Open-Domain Question Answering
with BERT

A natural language processing research paper analyzing the difference between reading comprehension and open-domain question answering. This paper examines two potential factors, context length and question generation, that make open-domain question answering the more difficult task. The Stanford Question Answering Dataset (SQuAD 2.0) is used as the baseline reading comprehension dataset and Google Natural Questions (NQ) as the open-domain dataset. A BERT-base uncased model is fine-tuned on the datasets to run experiments evaluating model performance.

• Research Paper

Predicting Flight Departure Delay

A machine learning project predicting flight departure delays using an airlines dataset of 30 million records and weather dataset of 630 million records. Using the Databricks platform, data preparation and feature engineering are performed on the two datasets. Categorical data is indexed, one hot encoded and assembled into vectors along with numerical data. A data pipeline is built in preparation for algorithm exploration. Using Apache Spark’s machine learning library, scalable logistic regression, decision tree, and random forest classification models are conducted.

• Report
• Presentation

Effects of Race on Perceptions of Professionalism

A causal inference experiment examining the effects of an individual’s skin tone on how others view their level of professionalism. The project is based on my research experiment proposal that was reviewed by academic peers and selected as one of the top 3 proposals to proceed with execution of the collaborative final experiment. The survey data is collected using Qualtrics and Prolific platforms. An analysis involving hypothesis testing and OLS regression for within-subjects experiment is conducted using R statistical software.

• Research Paper
• Experiment Proposal

House Price Prediction

A collaborative machine learning project to predict home sale prices using a housing dataset with 80 explanatory variables. Using Python and Jupyter Notebook, the project runs an ensemble of advanced regression techniques including linear regression, decisions trees with gradient boosting and neural networks. My individual contributions include the use of feature selection methods to refine linear regression models resulting in a mean log square error that was a 70% improvement over the baseline model.

• Notebook
• Presentation

Reducing Crime

A collaborative project analyzing a dataset of state crime statistics, which includes a dependent variable of crime rate and independent variables grouped into categories of deterrent, demographic, economic and geographical factors. The key determinants of crime are identified using R statistical software to build regression models to analyze the explanatory variables with practical and statistical significance. The results of the models are used to make policy recommendations directed at reducing crime levels.

• Report

Bike Share Query Project

A query project that uses the Lyft Bay Wheels dataset to gain insight into popular bike share commuter trips within San Francisco. Utilizing SQL and BigQuery within the Google Cloud Platform (GCP), queries are made from a Jupyter Notebook through a virtual machine and presents findings using Pandas output tables and Matplotlib visualizations. The data analysis results are used to answer business-driven questions and provide recommendations to increase ridership.

• Notebook

Analyzing Twitter Data for #DataScience

A collaborative Python project conducting a data analysis of Twitter users who include the hashtag #DataScience. The project uses the Tweepy package to access JSON structured data through the Twitter API. Using the Twitter data, an analysis determines the influential users, sentiment, and location of those who use the #DataScience. My individual contributions include an analysis of the location of twitter users with presentation of findings through visualizations using Matplotlib and Geopandas.

• Report
• Presentation