Slow But Flexible Or Fast But Rigid? Discrete And Continuous Processes Compared by Matteo Priorelli, Ivilin Peev Stioianov instant download

Slow But Flexible Or Fast But Rigid? Discrete And Continuous Processes Compared by Matteo Priorelli, Ivilin Peev Stioianov instant download

A tradeoff always exists when dealing with multi-step tasks. Higher-level processes can find theA tradeoff always exists when dealing with multi-step tasks. Higher-level processes can find thebest sequence of action primitives to achieve goals in uncertain environments, but they are slow andrequire significant computational demand. Contrarily, lower-level information processing allowsreacting to environmental stimuli rapidly, but the capacity to determine the optimal action when anenvironmental change occurs is limited. Through reiteration of the same task, biological organismsfind the optimal tradeoff: from primitive movements, composite actions gradually emerge by creatinglow-level task-specific neural structures. What are the underpinnings of such mechanisms and howto exploit them in robotics? We adopted the Active Inference perspective that casts behavior as aminimization of prediction errors, and compared two hierarchical strategies on a pick-and-place task:a discrete-continuous model with global planning capabilities and a continuous-only model withfixed transitions. We show that different phases of motor learning can be expressed in these terms,and propose how discrete actions might be encoded into continuous representations, which helpssolve the task fast, smoothly, and with low computational burden. Overall, our study paves the way tounderstanding task-specialization mechanisms and how they can be adopted in intelligent agents.1 IntroductionReal-world tasks generally involve complex actions consisting of multiple primitive movements. Even a simplepick-and-place operation consists of opening the hand while moving toward an object, grasping it, moving the armto a desired position, and finally releasing it. How does the brain support the efficient execution of such complexactions? While a sequence of movements could be easy to realize in static contexts, a difficulty emerges when actingin a dynamic environment, e.g., when a moving object has to be grasped on t