Heterogeneous-speed recurrent neural networks: Multi-timescale dynamics for fast and stable motion perception

Perception in dynamic environments demands both speed for detecting transients and stability for integrating noisy evidence over time. We propose a general dynamical framework that links multi-timescale, delay-coupled neural dynamics to an interpretable, trainable architecture for motion perception. Starting from extracellular recordings, neurons are stratified by spike width and firing rate into narrow-, intermediate-, and broad-spiking classes, revealing distinct latencies and durations to moving versus looming stimuli. These observations motivate a continuous-time rate model with heterogeneous time constants, explicit input/recurrent delays, and a cluster-level divisive normalization that bounds population activity. We further analyze sufficient conditions under which the delayed dynamics are well posed and bounded, and show how timescale separation gives rise to an interpretable fast/slow decomposition. Via delay-consistent discretization, the model yields a Heterogeneous-Speed RNN (HS-RNN) whose blocks map one-to-one to the mechanistic terms, enabling interpretable ablations. On motion-sequence classification and synthetic object-detection tasks, HS-RNN achieves superior performance to the compared recurrent models, whereas on static-sequence tasks its advantage is much less pronounced. Hidden states reproduce physiological fingerprints – shorter latencies and stronger transients for moving stimuli alongside longer sustained responses – demonstrating biological consistency. Although validated on the avian Entopallium-Mesopallium Ventrolaterale pathway, the framework is modality- and species-agnostic, offering a principled route from circuit dynamics to robust sequence models.