Erin Yoo
Pronouns: she/her
Research Mentor(s): Stefan Larson
Co-Presenter: Cade, Ciara
Research Mentor School/College/Department: SkySync / NonUM
Presentation Date: April 20
Presentation Type: Poster
Session: Session 5 – 3:40pm – 4:30 pm
Room: League Ballroom
Authors: Stefan Larson, Erin Yoo, Ciara Cade
Presenter: 75
Abstract
The development of artificial intelligence and machine learning gives computers the capability to automatically identify and categorize different files and images. When the process of identifying information and data becomes inconsistent, artificial intelligence software can become unreliable to automate specific tasks compared to humans. Our image detection model relies on machine learning, which uses training data to “fine-tune†and observe the inconsistencies to configure models for classification tasks. In order to ensure model robustness, the model predictions must be manually verified. This research project aims to formulate a process by which an image object detection model is verified on a large set of images. To create more reliable data and recognition software, we trained the program to identify images by various object distribution classes using approximately 1.2 million visually similar images and documents from internet databases. We flagged false positives by only looking at images where the model predicted an object with over 0.4 confidence, as our current preliminary estimates show that the model makes incorrect predictions roughly 4% of the time on random images sampled from the internet. The false positives given by the program were then analyzed to comprehend how it processes and identifies different images, and configures after human correction. False positives demonstrate the importance of checking the database to help us understand where the algorithm is having difficulty. Analyzing false positives also sets higher benchmarks for future identification software, decreasing room for error and becoming more efficient with automating the processing of classification. Familiarity with how artificial intelligence can become more robust at processing information and categorizing data by machine learning opens the door for automated technology and machines’ perception of their environment.
Engineering
