Implicit Design Choices and Their Impact on Emotion Recognition Model Development and Evaluation

Virtual Event: Zoom  Meeting ID: 920 6438 6553  Passcode: 753173

Abstract: In this dissertation, I focus on the subjective and variable nature of prediction and perception of emotion for any downstream task purposes. I explore various methodologies that can improve the usability and viability of the emotion recognition models in real world settings. In particular, this dissertation focuses on three major problem criteria that arise in the domain of emotion, but are also applicable to other domains. These criteria are: (a) ensuring that machine learning models are robust to variations and confounding factors, (b) obtaining the required training data in a cost-effective manner, and (c) using model interpretability and existing sociological literature to define zero-cost human-centered metrics for model training.