MSR-SNN: Advancing Neuromorphic Vision Object Classification With Multiscale-Residual Spiking Neural Networks

Spiking neural networks (SNNs), a class of brain-inspired computational models, are characterized by asynchrony, discreteness, and sparseness, attributes that are congruent with neuromorphic data. Recent research has been dedicated to applying SNNs to neuromorphic data classification tasks for enhancing classification performance. However, this task is significantly constrained by the expressive abilities of neurons and the efficient feature extraction technology for spiking signals. Motivated by these, we first propose the generalized parameter leaky integrate-and-fire (G-PLIF) neuron model, in which neuron membrane potential threshold is variable to simulate neurons with different biological characteristics. In G-PLIF, the neuron membrane potential threshold is constantly changing during the training process, thereby enhancing the dynamic heterogeneity and representational ability of neurons. Then, based on G-PLIF, we designed a multiscale-residual SNN (MSR-SNN) specifically for the classification tasks of neuromorphic vision objects. The concept of a multiscale pyramid has been integrated into the residual blocks of the network, fortifying its ability to extract deep-level, multidimensional features. The proposed method is evaluated on challenging neuromorphic datasets dynamic vision sensors (DVS)-CIFAR10 and DVS-Lip, and extensive experimental results demonstrate that our method achieved state-of-the-art performance.