Abstract:
Studying freely moving animals is essential for understanding natural behaviors such as locomotion, foraging, or social interactions. Recent advancements in cameras, motion capture, and pose estimation have enabled high-throughput analysis of animal movement. While deep learning (DL) methods are widely used for human movement analysis and sequential data, their application to animal behavior remains is still emerging, primarily for pose estimation. Motion capture and pose estimation provide high-quality data on individual movement but often contain missing values that hinder downstream analysis. To address this, I developed Deep Imputation for Skeleton Data (DISK), a deep learning algorithm that reconstructs missing tracking data by leveraging spatial and temporal dependencies between keypoints. We demonstrated its effectiveness across species and behaviors and developed it into a user-friendly tool with an uncertainty score to assess imputation quality (github.com/bozeklab/DISK.git). My research focuses on developing flexible DL methods to link animal movement with neural activity, utilizing unsupervised and transfer learning approaches to minimize manual labeling. I aim to investigate how biological perturbations—genetic, pharmacological, and social—modulate behavior, ultimately creating a framework to enhance our understanding of behavioral dynamics. A precise, data-driven description of natural, unconstrained behavior could also inform the diagnosis and treatment of musculoskeletal and neurological disorders.