Temporal-context-aware dynamic encoding with a VO2 memristive neuron

Abstract Efficient dynamic perception requires sensory systems to capture not only instantaneous stimulus changes but also the temporal context established by successive events. Here, we demonstrate a temporal-context-aware dynamic encoding scheme based on a volatile VO2 memristive neuron. By progressively extending conventional leaky integrate-and-fire operation to phasic and short-term-plasticity-inspired dynamics, the proposed neuron transforms sustained inputs into sparse transient spikes while incorporating recent event history into spike generation. Rising and falling input transitions are selectively encoded as transient spike responses, and a decaying internal state further modulates subsequent responses according to event polarity and inter-event timing. Successive events with the same polarity enhance the subsequent response, whereas those with opposite polarities suppress it, with the modulation strength continuously governed by the inter-event interval. Consequently, event polarity, temporal proximity, and recent stimulus history are jointly embedded into spike generation at the sensory front end, reducing the need for downstream extraction of temporal correlations. Integrated with a spiking neural network, the proposed encoding scheme achieves a mean classification accuracy of 89.06% on a four-class UTD-MHAD action-recognition task, substantially outperforming conventional leaky integrate-and-fire and phasic leaky integrate-and-fire encodings. These results demonstrate that coupled device–circuit dynamics can provide a compact hardware substrate for context-dependent spike encoding and offer a promising route toward efficient neuromorphic dynamic perception.