Date of Award:

8-2020

Document Type:

Thesis

Degree Name:

Master of Science (MS)

Department:

Computer Science

Committee Chair(s)

Douglas Galarus

Committee

Douglas Galarus

Committee

John Edwards

Committee

Nicholas Flann

Abstract

We are witnessing an influx of data - images, texts, video, etc. Their high dimensionality and large volume make it challenging to apply machine learning to obtain actionable insight. This thesis explores several aspects pertaining to dimensional reduction: dimension reduction methods, metrics to measure distortion, image preprocessing, etc. Faster training and inference time on reduced data and smaller models which can be deployed on commodity hardware are a critical advantage of dimension reduction. For this study, classical machine learning methods were explored owing to their solid mathematical foundation and interpretability.

The dataset used is a time series of images from several camera feeds observing the traffic, weather and road conditions along highways. The time-series nature of dataset gives rise to interesting questions which are investigated in this work. For instance, can machine learning models trained on past data be used on future camera feed data? This is highly desirable and yet difficult due to the changing weather, road conditions, traffic conditions and scenery. Can dimension reduction models obtained from past data be used for reducing dimensionality of future data? This thesis also examines the difference between the performance of machine learning methods before and after application of dimension reduction. It tests some existing metrics to measure quality of dimension-reduced data set and introduces several new ones. It also examines the application of image pre-processing methods to boost the performance of classifiers. The classification performance with and without random sampling has been studied as well.

Checksum

ebfd254650ee831dcb68e49310929992

Share

COinS