Simulated Learning of Rhythm Perception and Synchronization
Co-Authors:Yassaman Ommi, Manav Shardha, Jonathan Cannon
Virtual or In-person:In-person
Sensorimotor synchronization, the coordination of rhythmic auditory input and motor responses, plays a crucial role in musical engagement and social bonding. Humans often show spontaneous audiomotor synchronization to rhythms, while most animals do not. Recent experiments, however, reveal the remarkable capacity of reinforcement to teach synchronization with metronomes to non-inherently synchronizing animals, prompting the question of whether infants acquire this skill similarly. This study investigates the neural processes underpinning rhythm perception and beat prediction while exploring methods to train artificial agents to entrain to rhythms. First, we used backpropagation to train a Recurrent Neural Network (RNN) to anticipate a beat by generating sine waves at a tempo matching a train of input pulses. Focusing in on the skill of timely prediction, we used a target-based algorithm called Full-Force, to train an agent to generate predictive pulses immediately preceding each event in an isochronous train of input pulses.
Next, considering the basal ganglia's role in rhythm perception and reinforcement, we have begun development of a Reinforcement Learning (RL) agent with the eventual goal of teaching it to synchronize rhythmic movements with a rhythmic input using a similar reinforcement scheme to that used to train animals to synchronize. Employing RL techniques, we effectively trained the agent to adjust its tapping tempo in response to contextual cues. Our goal is to explore training methods for artificial agents and networks in sensorimotor synchronization tasks, integrating our computational findings with neural data from animal and human experiments to explain the learning process.