Name:Yassaman, Ommi

School/Affiliation:McMaster University

Co-Authors:Matin Yousefabadi, Jonathan Cannon

Virtual or In-person:In-person

Abstract:

Sensorimotor synchronization (SMS) is the coordination of movement with external rhythmic stimuli, such as tapping your foot along to a beat. Humans learn to synchronize over the course of development, but it is not clear how this learning is driven or motivated; monkeys, too, can learn to synchronize, but only when rewarded according to carefully designed policies. We aimed to create a model of learning to synchronize through reward in order to provide a new fount of insight into this process in humans and animals. We developed a recurrent neural network (an "agent") that could sense "clicks" and produce "taps," and trained it using deep reinforcement learning. We trained the agent on rhythmic tapping tasks, simulating the synchronization behavior observed in human and monkey studies. Different reward strategies were applied, focusing on minimizing tap asynchrony, improving inter-tap interval accuracy, and encouraging anticipatory tapping over reactive tapping. The performance was further evaluated in perturbation tasks where timing shifts were introduced. Agents trained with combined rewards exhibited human-like synchronization patterns, such as asymmetric error correction, adjusting more after tapping late than after tapping early. We conclude that an appropriate policy of reward-driven learning can account for specific adaptive rhythmic behavior in humans, which may emerge from the structure of rewards rather than specific neural architectures. These findings, and further exploration of the model, promise to provide a new perspective on the developmental and neural underpinnings of human timing abilities, and may have implications for developing AI systems capable of human-like rhythmic coordination.