Dendritic functional heterogeneity with feedback control facilitates stable spatiotemporal representation in spiking...

Stable and accurate spatiotemporal representation is crucial for spiking neural networks (SNNs) to perform well in modeling long sequences. Existing single-compartment models struggle to capture long-term dependencies, while multi-compartment models often neglect the functional complementarity and predictive interactions among dendritic subunits. Both models limit the spatiotemporal representation capacity of SNNs. To overcome these limitations, we propose the Dendritic Feedback Adaptive Leaky Integrate-and-Fire (DF-ALIF) neuron model, which combines dendritic functional heterogeneity with feedback control within a predictive coding framework. Basal compartments of DF-ALIF receive feedforward and recurrent inputs, encoding bottom-up prediction errors via a local excitatory and inhibitory balance mechanism that drives feature detection and pattern reconstruction. In contrast, the apical compartment carries top-down feedback predictions and dynamically modulates the somatic firing threshold via a mismatch mechanism to calibrate neural activity. The synergistic interaction between these compartments mitigates the temporal credit assignment problem and enhances the spatiotemporal representation capabilities of SNNs. On benchmarks including electrocardiogram classification, radar and neuromorphic event-based gesture recognition, image classification, and speech recognition, DF-ALIF demonstrates improved classification accuracy, stable convergence, and enhanced robustness to temporal perturbations. Our work demonstrates a feasible approach to embedding dendritic functional heterogeneity in SNNs and offers a fresh perspective on the design of neuromorphic computing systems.