Jonathan McDevitt Jr
Pronouns: he/him
Research Mentor(s): Ada Eban-Rothschild
Research Mentor School/College/Department: Psychology / LSA
Program: UROP
Session: Session 1 (9:00am – 9:50am)
Authors: Jonathan McDevitt, Gaurav Kaul, Ada Eban-Rothschild
Abstract
This project explores automatically curating an image dataset for training machine learning (ML) models from a large repository of animal behavior videos. A basic approach to this is to evenly or randomly sample frames from the videos, but this may result in the dataset containing redundancies, such as multiple images where the animals are sleeping or simply remaining in the same position. Due to the size of this dataset, this would require additional and unnecessary manual image labeling and ML model training time. For this reason, this project searches for a method of reducing the size of this image dataset by removing these redundancies. In this work we present an automatic curation process that functions by computing “distances†between the images in a dataset, and removing the images that are closest to other images. This project explores various metrics for computing distance, including random, Euclidean, cosine, color distribution, and neural network-encoded distance metrics. Each of these metrics are used to curate image datasets, and an ML model is trained on each of the datasets separately under otherwise identical training conditions to complete tasks related to animal behavior analysis. The resulting performance on animal behavior classification is used to evaluate the distance metrics’ effectiveness. Although this project focuses on animal behavior data, it also explores generalizing this method to other image datasets, including MNIST and CFAR-10.
