Fully exploiting temporal dynamics in spiking neural networks

Determine how to fully exploit temporal computational principles in spiking neural networks—including optimization and utilization of neuronal membrane time constants, synaptic delays or causal temporal convolution parameters, adaptive threshold mechanisms, and recurrent connectivity—to achieve maximal performance and parameter efficiency on temporal processing tasks.

Background

Spiking neural networks (SNNs) leverage temporal dynamics for computation, and recent work has shown that optimizing elements such as neuronal time constants, synaptic delays or causal temporal convolutions, adaptive thresholds, and recurrent connections can improve accuracy and reduce parameter counts on temporal tasks.

Despite these advances, the authors explicitly note that fully realizing the potential of these temporal computational principles remains unresolved, motivating their investigation of temporal hierarchy as an inductive bias while acknowledging that the broader question persists.