Abstract/Summary Tracking the movements of objects and parts of objects - referred to as pose estimation - is critical for understanding the mechanisms underlying complex behavior. Characterizing dynamic behaviors of animals (and other systems) is central to many disciplines, including computer science, physics, ethology, kinesiology, and sports medicine. Here we focus on neuroscience, where linking brain activity to associated dynamic behaviors is critical for both understanding normal function as well as effects of injury, disease, or degeneration. Invasive methods for measuring behavior are highly accurate, but require placement of sensors that may themselves interact with behavior and which may be susceptible to deterioration or infection. Video provides a non-invasive approach to characterizing behavior over time. Extracting behavior from video streams has, historically, been a slow and laborious process. Recent work in machine learning and artificial neural networks (ANNs), though, has revolutionized this process, making the analysis of complex video far easier and more accurate. While these systems are highly flexible, they were not designed for real time use, meaning that large video files must first be stored to disk for subsequent analysis. This poses two problems that this proposal will attempt to address. First, there is significant cost and management challenges associated with storage of large video stores, forming a practical barrier for adoption of this important technology for characterizing behavior. Second, estimates related to behavioral state are not available in real time so they cannot be used to control the experiment. We will develop a research methodology for “closing the loop”, by taking the networks trained by an existing and highly successful markerless object tracking system (DeepLabCut) and optimizing them for real time inference. After the system is functional, verified, and benchmarked, it will be shared with the community through open source repositories.