Predictive Power of Behavior: Robust Low-Dimensional Neural Representations from Pose Data
1 UKB
2 University of Bonn
Understanding neural activity is central to neuroscience. While high-bandwidth data acquisition techniques exist, they often rely on invasive methods that restrict the free movement of animals and raise ethical concerns. Recent advancements in markerless pose estimation enable the quantitative characterization of animal behavior. Since behavior is closely linked to neural activity, this approach can serve as a non-invasive proxy for brain recordings. However, the high-dimensional data generated by these methods can obscure underlying patterns. Despite their size, these datasets are hypothesized to conceal a low-dimensional representation that correlates behavior with neural states. Recent developments in machine learning have made it possible to extract and uncover this low-dimensional representation from complex, high-dimensional data. In this study, we evaluate the robustness of these representations by predicting future behavioral information. Additionally, we demonstrate the practical application of our framework in early seizure detection based solely on pose recordings.