B-SOiD Automated Quantification of Naturalistic Behaviors with Supervised or Unsupervised Approaches

To make movement and foraging decisions in a naturalistic environment, multiple neural populations must work synergistically to produce successful actions. These decisions span multiple scales, from kinematic control of vigor, to broader decision motifs associated with a given context. We have recently developed a suite of tools for rapid, accurate behavioral segmentation that now allows for the identification and interrogation of spontaneous, naturalistic behaviors. This suite includes both unsupervised (B-SOiD) as well as supervised algorithms (A-SOiD) which learn behaviors and add interpretability to pose estimation data. We simultaneously recorded from neurons across motor cortex and striatum as a mouse freely explored an open-field over the course of several days. We then these techniques to identify distinct movements, behaviors, and action motifs. B-SOiD identified significant patterns of co-activation across neurons in these areas. Surprisingly, all of the identified behaviors elicited a robust neural signature, though we note several differences in the sparsity and dimensionality of the representation in different areas. A simple decoder based upon individual single-neuron spike rates was able to predict the behavior performed at a level many times greater than chance, demonstrating the strength of the neurobehavioral relationships. This strength also enables the responses of single cells to be well-explained by a small set of underlying ‘latent’ dimensions. Traditional methods involving linear tools (e.g. PCA, GPFA) have been successful in explaining shared variability in neural populations during artificial, overtrained tasks; however, how populations covary in naturalistic settings absent reductionist tasks and rigid restrictions on movement has yet to be evaluated. We show that the naturalistic behaviors of mice in the open field require an appreciation of more complex dimensionality. Specifically, we observed that many latent dimensions were nonlinearly related to recorded neural firing rates, particularly in dorsal striatum. These results, including the differing levels of complexity underlying the representation of behavioral features, shed new light into how the coordinated translation of information between areas orchestrate complex, ethological behaviors.

Learning Objectives: 

1. Understand the difference in substance and application between supervised and unsupervised machine learning algorithms.

2. Determine a baseline understanding of how to apply the B-SOiD behavioral segmentation platform to study a variety of behaviors and datasets.

3. Identify basic differences in how motor cortex and basal ganglia represent spontaneous, naturalistic behavior.