Automated Phenotyping of Mouse Social Behavior

Inspired by the connections between social behavior and intelligence, I have developed a trainable system to phenotype mouse social behavior. This system is of immediate interest to researchers studying mouse models of social disorders such as depression or autism. Mice studies provide a controlled environment to begin exploring the questions of how to best quantify social behavior. For the purposes of evaluating this system and to encourage further research, I introduce a new video dataset annotated with five social behaviors: nose-to-nose sniffing, nose-to-head sniffing, nose-to-anogenital sniffing, crawl under / crawl over, and upright head contact. These four behaviors are of particular importance to researchers characterizing mouse social avoidance [9]. To effectively phenotype mouse social behavior, the system incorporates a novel mice tracker, and modules to represent and to classify social behavior. The mice tracker addresses the challenging computer vision problem of tracking two identical, highly deformable mice through complex occlusions. The tracker maintains an ellipse model of both mice and leverages motion cues and shape priors to maintain tracks during occlusions. Using these tracks, the classification system represents behavior with 14 spatial features characterizing relative position, relative motion, and shape. A regularized least squares (RLS) classifier, trained over representative instances of each behavior, classifies the behavior present in each frame. This system demonstrates the enormous potential for building automated systems to quantitatively study mouse social behavior. Thesis Supervisor: Tomaso Poggio Title: Eugene McDermott Professor