Spiking neuron model learns sequence timing with oscillatory clock control

A team led by Melissa Lober from the research consortium (including Bouhadjar, Diesmann, and Tetzlaff) has published a preprint on arXiv extending the spiking Temporal Memory (sTM) model to learn both the order and exact timing of sequence elements. In its original form, the sTM model could encode the identity of sequence elements in context but lacked a mechanism for element-specific duration. The new work introduces a population-based encoding where each element's duration is represented by the sequential activation of dedicated neuronal groups, allowing the network to represent complex temporal patterns across a wide range of timescales.

Crucially, the study demonstrates that oscillatory background inputs—similar to brain rhythms observed in EEG or LFP recordings—can act as a global clock signal. By modulating the frequency of these oscillations, the network can flexibly control the speed of sequence replay, mimicking how the brain replays memories at different speeds during wakefulness and sleep. The authors show that elapsed time is encoded by unique, sparse spatiotemporal activity patterns. This biologically plausible framework bridges computational neuroscience and cognitive function, offering a potential substrate for sequence learning in neural circuits and inspiring new approaches in neuromorphic hardware design.