Date: Friday, November 8, 2024

Time: 3:00 p.m. - 4:00 p.m.

Location: Centergy Building 5126. The associated zoom link is: https://gatech.zoom.us/j/97484699776 

Speaker: Chen Zhou

Speakers' Title: Ph.D. candidate in the Omni Lab for Intelligent Visual Engineering and Science at Georgia Tech

Seminar Title: Learning from Expertise-based Annotation Disagreement

Abstract: Label disagreement describes the subjectivity of multiple annotators interpreting the same data. Traditionally, disagreement is considered to hinder the development of models that rely on gold-standard training labels. Resolving disagreement is more challenging in the fields where obtaining gold-standard labels require a certain level of expertise. In this talk, I will discuss the characterization of expertise-based label disagreement in subsurface fault segmentation, and empirically demonstrate that incorporating disagreement labels can enhance the performance of neural networks with limited gold-standard labels. In particular, I will first analyze the behavior of disagreement modeling under two suboptimal assumptions of labels. Based on the observations, I will derive an effective disagreement characterization with the proper assumption of expertise-based labels. Furthermore, I will integrate this approach into a model training workflow to enhance the performance with a mixture of annotations from two different levels of expertise. The talk concludes with the insight that properly allocating imperfect labeling on large amounts of targeted data can be beneficial in developing models with limited access to gold-standard labels.

Bio: Chen Zhou is a Ph.D. candidate in the Omni Lab for Intelligent Visual Engineering and Science (OLIVES) at the Georgia Institute of Technology, working with Prof. Ghassan AlRegib. His research interest includes generative modeling, uncertainty quantification, and label-constrained learning across multiple visual applications including image/video understanding and computational image interpretation.