Hierarchical Control of Reaching Movements Via Compositional Gain Modulation

Reusing existing modules in novel settings via compositional generalization is the hallmark of intelligent behavior. While much research is dedicated to studying how to enable AI systems to learn and reuse modules effectively for better performance and increased computational efficiency, there is still a lack of consensus on which modules the brain leverages and how to identify them. To shed some light on this matter, here we investigate the modularity principles the brain uses to control the body efficiently. After briefly revisiting established models of domain-specific spatial and temporal motor modularity, we introduce a new, unifying computational model of compositional generalization in the motor system based on the Canonical Polyadic Decomposition (CPD) model. We show that the model --- which leverages gain modulation --- can simultaneously capture modularity in the spatial, temporal, and action domains with a lower number of parameters than established models. Furthermore, we show that the geometrical organization of the action modules the model isolates is not random but describes a smooth manifold that allows the zero-shot learning of muscle patterns for untrained movements. Taken together, our results suggest that the decomposition proposed here represents an effective compositional strategy the brain could leverage to control complex movements while saving computational resources.