Study of the Effect of Customized Training Sets in AI Emotion Detection

Presenter

Varun Gopal

Campus

UMass Amherst

Sponsor

Steven D. Brewer, Department of Biology, UMass Amherst

Schedule

Session 2, 11:30 AM - 12:15 PM [Schedule by Time][Poster Grid for Time/Location]

Location

Poster Board A10, Campus Center Auditorium, Row 1 (A1-A20) [Poster Location Map]

Abstract

Significant advances in machine learning technology have led to the development of multiple AI-powered algorithms that can identify emotions of subjects based on their facial expressions. These algorithms use large, standardized, and labeled sets of image data to “teach” the AI to accurately identify the emotion of any subject given to the algorithm. Within these algorithms, opinions are divided regarding the accuracy of the algorithm’s emotional recognition when used with diverse populations, with some finding a decrease in emotional identification accuracy when used with non-white, non-male and not middle aged subjects. Here I attempt to resolve this bias found in present algorithms, by custom training the algorithms to individuals by creating a tagged set of data customized to a subject and their unique emotional expression. Here I describe my findings when comparing the accuracy of a custom-trained emotional detection algorithm, when compared to a control algorithm, trained on a standardized training set, akin to how current algorithms are trained. Accuracy of the algorithm was assessed by observing algorithmic identification when compared to surveying of subjects on their perceived emotion. Addressing bias in facial recognition algorithms paves the way for this style of algorithm to begin being used with more diverse populations, such as in the monitoring of psychiatric patients, or in aiding subjects who struggle to convey their emotions to others.

Keywords

Emotion Detection, Machine Learning Algorithms, Bias

Research Area

Artificial Intelligence

SIMILAR ABSTRACTS (BY KEYWORD)