For decades, neuroscience has focused almost exclusively on stereotyped, reductionist, and over-trained behaviors due to their ease of study. In contrast, naturalistic behavior provides a rich diversity of movements, but this feature also largely precludes it from quantification and use. Recent advances in computer vision have enabled automatic tracking of the position of body parts – but position is not behavior. To provide a bridge from positions to behaviors and their kinematics, we developed B-SOiD (Hsu and Yttri, Nature Communications). This open-source method discovers natural spatiotemporal patterns in body position data, then uses the cluster statistics to train a machine learning algorithm to classify behaviors that can generalize across subjects and labs. We will discuss the application of this user-friendly algorithm in flies, mice, and humans. Finally, we will share new data from recordings throughout the cortex and basal ganglia that reveal how these diverse behaviors are encoded by single units and interconnected neural populations.
Zoom option if unable to attend in person