McMaster University
Emergent Rhythmic Timing in Generic Recurrent Neural Networks Trained to Predict
Abstract
dc:description.abstractMoving in time with a beat demands anticipation: a movement timed to a moment that has not yet happened. Whether this rests on circuitry evolved for the purpose or emerges from more general machinery shaped by experience remains open. This dissertation asks whether human-like rhythm perception, anticipation and synchronization arise in generic recurrent neural networks trained only to predict, with no oscillator, beat detector or clock built in. Three networks put the question in three forms. A continuous-time network trained only to anticipate isochronous pulses reproduced the behavioural signatures of human timekeeping, including Weber's law, a bias toward the mean of the trained tempi, and overshoot after a tempo change, and organised itself into the low-dimensional regenerating trajectory and sequential single-cell code of primate medial premotor cortex. A predictive-coding network learning by a local Hebbian rule acquired a periodic internal dynamic, and the anticipation of a periodic sound, from nothing but the correlated auditory and vestibular stimulation of maternal gait; controls show that the vestibular signal's continuity permits the dynamic, and its correlation with sound gives the auditory channel access. A long short-term memory network with its output returned to its own input went on producing a beat after the beat stopped arriving; where that signal was routed decided whether continuation occurred at all. The claim is sufficiency, not necessity. These signatures are obtainable with no timekeeping apparatus supplied, reframing them as what anticipation looks like in any sufficiently expressive recurrent system, not as evidence of circuitry evolved for one. A disagreement with the primate recordings proves more instructive: networks that continue indefinitely discard the beat-number information that monkey cortex retains, while those that stop preserve it. What a timing system represents depends sensitively on what its task requires, in animals as much as in models.
Degree
thesis:*- Department dc:contributor.department
- Computational Engineering and Science
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Yousefabadi, Matin
- Advisor dc:contributor.advisor
-
- Cannon, Jonathan
Subjects
dc:subject × 7Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Identifier URI
- https://doi.org/10.71548/25584