Myra Lyu
Research Mentor: Bing Ye
Mentor Department: Life Sciences Institute, University of Michigan, Medicine
Author(s): Myra Lyu, Bobby Tomlinson, Bing Ye
Session: Session 4 (1:00 PM – 1:50 PM)
Presentation Type: Poster 51
Abstract
Automated animal behavior analysis has become increasingly important in neuroscience, behavioral biology, and pharmacological research. LabGym is an open-source tool that uses machine learning to quantify animal behaviors from video data and provides a variety of outputs to support downstream analysis. However, the current outputs are often dense and difficult for users to interpret efficiently, limiting the accessibility of model predictions and the ability to evaluate behavioral dynamics. This project investigates methods to improve the interpretability and usability of LabGym’s behavioral analysis outputs. I examined the behavior-analysis module within the LabGym codebase, evaluated its existing outputs, and compared its functionality with that of related animal behavior analysis tools. Based on this evaluation, I identified two features that could significantly enhance interpretability and comparability of behavioral predictions: a frame-level probability matrix and a state transition map. The frame-level probability matrix records the probability distribution of predicted behaviors for each frame in a video, allowing researchers to assess model confidence and identify ambiguous or transitional behaviors. The state transition map summarizes the temporal relationships between behaviors, providing a visual representation of behavioral dynamics over time. Together, these proposed features provide a more transparent representation of model predictions and behavioral patterns, facilitating deeper insight into behavioral datasets. The results of this study provide a foundation for improving LabGym’s analysis modules and aim to enhance the accessibility, interpretability, and analytical power of AI-based tools for animal behavior research across multiple scientific disciplines.
