Learning to synchronize: A delay-based plasticity rule for temporal coding in spiking networks

The traditional view of neural network learning focuses on classification of static patterns. For spiking neural networks processing continuous dynamic signals, however, a more fundamental task may be the temporal coordination of incoming information. This work introduces the Synchronizing Learning Rule (SLR), an unsupervised plasticity rule for spiking networks. Unlike conventional approaches, SLR does not modify synaptic weights. Instead, it directly regulates signal conduction delays using local spike-timing-dependent traces, synchronizing presynaptic spike arrivals at the postsynaptic neuron to maximize its response to a repeating temporal pattern. In a temporal pattern recognition task, SLR-trained networks demonstrate two key advantages over weight-based STDP rules. They operate reliably across a substantially wider range of parameters and, crucially, exhibit superior scalability — successfully learning patterns nearly twice as long as those learnable by weight-based methods. These results establish direct temporal synchronization as an efficient mechanism for processing dynamic data in spiking neural networks.