Neuromorphic systems offer energy-efficient solutions for temporal signal processing by emulating the dynamics and heterogeneity of biological neural circuits. However, conventional approaches face challenges in adaptive regulation and in capturing multi-timescale temporal features. Here, we present a bio-inspired neuromorphic hardware system that integrates homeostatic neurons with programmable dendritic structures. Utilizing the threshold-switching characteristics of VO2, we construct a homeostatic neuron enabling autonomous stabilization of neuronal activity. The dendritic module, co-designed using CMOS-RRAM and VO2 devices at the board level, enables programmable spike delays for multi-timescale temporal feature extraction. When embedded into a spiking neural network, the system achieves classification accuracies of 92.14% ± 0.99% for industrial defect detection and 86.53% ± 0.18% for speech recognition, while operating at 19.29 pJ per spike, surpassing conventional processors. The results demonstrate a scalable and biologically inspired hardware framework for efficient temporal signal processing, suggesting potential in next-generation neuromorphic accelerators. Current neuromorphic systems are constrained in handling temporally complex computation. Zhang et al. present a fully memristor-based neuron-dendritic system that endows multi-timescale signal processing capability, delivering accurate and energy-efficient computational performance.
Zhang et al. (Thu,) studied this question.