Association for Behavior Analysis International

The rapid advancement of computer technology, artificial intelligence (AI), machine learning, deep learning, and machine vision has created unprecedented opportunities in the study of animal and human behavior. These tools enhance the precision and speed, of behavior data collection and analysis. AI-based systems, particularly those using machine learning and deep learning, automate complex tasks like behavior recognition and movement analysis, reducing human error. Machine vision and pose estimation provide detailed insights into the spatiotemporal dynamics of behavior, allowing for a deeper understanding of both individual actions and broader patterns. This symposium presents four studies demonstrating cutting-edge AI applications in behavior analysis. The first applies real-time tracking, machine vision, and a closed-loop system to develop spatial contingent schedules of reinforcement. The second uses pose estimation to analyze and identify behavioral schedules in rats. The third study compares the accuracy of human observers and AI systems in recording gross motor movements and topography. The fourth introduces machine learning models that automatically measure vocal stereotypy in children with autism and examines the correlation between values measured by machine learning and those recorded by a human observer. Together, these studies show AI’s role in advancing research and applied practices in behavior analysis.

Audience members interested in artificial intelligence applications in the science of behavior will benefit from a basic understanding of computer-related terminology. This includes those seeking insights into integrating AI technologies for enhanced data collection and analysis in behavioral research and practice.

The Computational Experimental Analysis of Behavior (CEAB) is an emerging field that implements Computational Intelligence in methods and data analysis to understand behavioral phenomena. We implemented a novel Reinforcement Schedule (RS), based on real-time tracking with Machine Vision and a closed-loop system, called Spatial Contingent Schedules (SCS), in which reinforcement depends on a given spatial feature, i.e., a reached traveled distance. Two experiments were conducted under an A-B-A-C-A-B design, where A = Extinction, B = SCS, and C = JokedRS. We used a Modified Open Field System (1x1 m) with a servo water-dispenser at the center (Coord. .45, .45). In experiment 1, a Fixed Traveled Distance Schedule (FD) was used as SCS, and in experiment 2, a Variable Traveled Distance Schedule (VD). A Spatial Dynamics Behavioral Analysis, using Machine Learning, was conducted. The rats were sensitive to the contingencies for both SCS, and a variable ranking analysis showed that Spatial Anticipation of Reinforcement (SA) and Giving Up Time of the Reinforcement Zone were the most sensitive variables in distinguishing the effects between SCS. It is highlighted that SA was systematically observed under FD. The implications of CEAB and SCS for pushing the methodological and empirical boundaries of EAB are discussed.

Traditional behavior research often focuses on discrete responses, but organisms exhibit a wide range of behaviors in natural settings (Skinner, 1966). Spatiotemporal features of behavior are sensitive to reinforcement contingencies (León et al., 2020). Machine learning, both supervised and unsupervised, can detect behavior patterns in uncontrolled environments (Turgeon & Lanovaz, 2020). This study extends machine learning applications to identify behavior schedules in 12 rats using spatiotemporal data. Twelve three-month-old Wistar rats, individually housed with a 23-hour water restriction, were divided into four groups. They were exposed to fixed time (FT) and variable time (VT) schedules with fixed (FS) or variable space (VS) for water delivery over 30 sessions and 10 sessions without a programmed schedule. We utilized DeepLabCut for pose estimation, refining a pre-trained mouse model with our data. The primary measure was joint positions and skeleton representations from DeepLabCut. SimBA provided additional behavior recognition, and CEBRA offered latent behavior analysis. Results include reduced movement and behaviors under variable space and time contingencies due to environmental variability. This study implies that advanced machine learning can revolutionize behavior analysis, offering a more accurate understanding of animal and human behaviors, thereby better meeting species-specific needs in different natural environments.

Behavior frequency data is collected for individuals with autism to monitor progress and inform decision making by behavior analysts. This data impacts important life decisions, such as restrictive procedures fading, psychotropic medication prescribing, and residential placement (Vollmer et al. 2008). Studies have demonstrated exorbitant training and monitoring hours are required from a behavior analyst to ensure this data collected by direct care staff is accurate (Mozingo et al. 2006; Reis et al. 2013). The current comparative analysis evaluates the effects of human and AI data collectors on the accuracy of human gross motor frequency and topography data collection. Participants to date have consisted of typically developing adults who are prompted to make gross motor movements at latencies which decrease across sessions, and the topography and frequency of the movements are scored by a human data collector, a motion detection camera, and a generative AI system. Preliminary data suggest that the Axis camera system’s ability to discriminate frequency of behaviors decreases as the latency between movements shortens.