Enhancing learning efficiency in spiking neural networks via expanded representation and neuronal heterogeneity

Spiking Neural Networks (SNNs) remain limited by three coupled deficiencies in existing neuron models: binary spikes collapse continuous membrane states into 0/1 events, causing severe quantization loss; fixed or homogeneous thresholds cannot adapt to instantaneous inputs and neuron-wise dynamics; and conventional hard/soft reset rules either discard useful residual membrane information or retain excessive residual potential, leading to redundant firing. To address these limitations, we propose the dynamic threshold feedback reset leaky integrate-and-multi-valued-fire (DTFR-LIMF) neuron. Multi-valued firing expands spike states to mitigate spatial-domain entropy loss, an input-dependent dynamic threshold regulates firing difficulty at the neuron level, and a feedback reset mechanism applies history-dependent reward-penalty adjustment to improve temporal dynamics and suppress redundant spikes. For stable optimization of multi-valued firing, we further design a threshold-dependent surrogate gradient (TdSG) and an accelerated variant ( T 2 TdSG). To preserve event-driven efficiency during inference, we introduce an integer-to-binary (I2B) conversion scheme and binary-based energy metrics. Experiments on static and neuromorphic benchmarks show state-of-the-art performance, achieving 82.8% on CIFAR100, 69.89% on Tiny-ImageNet, and 89.4% on CIFAR10-DVS, while I2B reduces synaptic operations by 21%–36%.