Schedule of Events | Search Abstracts | Symposia | Invited Symposia | Poster Sessions | Data Blitz Sessions

Poster F123

Rhythmic Timing in Continuous-time Recurrent Neural Networks

Poster Session F - Tuesday, April 16, 2024, 8:00 – 10:00 am EDT, Sheraton Hall ABC

Manav Shardha1, Matin Yousefabadi1, Jonathan Cannon1; 1McMaster University

Humans' ability to anticipate rhythmic sequences across different tempos plays a crucial role in synchronization and music-making. Various animal species can learn to distinguish between isochronous and irregular rhythms, suggesting that the capacity for tempo-flexible rhythmic timing can be learned without the specialized mechanisms found in the human brains. To study the mechanisms underlying learnable rhythm perception, we studied solutions produced by continuous-time recurrent neural networks (ctRNNs), trained using full-FORCE or backpropagation. The ctRNNs were presented with an isochronous train of pulses across several tempos and were tasked with generating its own predictive pulses anticipating the input pulses. The full-FORCE-trained network exhibited a generalized mechanism for rhythmic prediction, accurately anticipating subsequent pulses and demonstrating increased precision with additional input pulses. This network reproduced key aspects of human timing tasks, displaying enhanced accuracy and reduced bias towards the mean tempo with longer trains of input pulses. It mirrored human-like adherence to "Weber's Law," with increases in pulse width and standard deviation of inter-pulse intervals proportional to increasing input period. The full-FORCE-trained network displayed circular oscillatory dynamics in neural phase space, resembling the neuronal data observed in monkey medial frontal cortices during synchronization tasks. These results could not be replicated in the backpropagation-trained network. The consistency of the full-FORCE-trained network's behaviour with human psychophysics results underscores the potential for achieving human-like perception and anticipation of isochronous sequences through semi-supervised (predictive) learning in generic recurrent networks without specialized timing mechanisms. Future research should focus on networks' capacity for more complex rhythm perception tasks.

Topic Area: PERCEPTION & ACTION: Other


CNS Account Login


April 13–16  |  2024