Deep CovDenseSNN: A hierarchical event-driven dynamic framework with spiking neurons in noisy environment

Abstract Neurons in the brain use an event signal, termed spike, encode temporal information for neural computation. Spiking neural networks (SNNs) take this advantage to serve as biological relevant models. However, the effective encoding of sensory information and also its integration with downstream neurons of SNNs are limited by the current shallow structures and learning algorithms. To tackle this limitation, this paper proposes a novel hybrid framework combining the feature learning ability of continuous-valued convolutional neural networks (CNNs) and SNNs, named deep CovDenseSNN, such that SNNs can make use of feature extraction ability of CNNs during the encoding stage, but still process features with unsupervised learning rule of spiking neurons. We evaluate them on MNIST and its variations to show that our model can extract and transmit more important information than existing models, especially for anti-noise ability in the noisy environment. The proposed architecture provides efficient ways to perform feature representation and recognition in a consistent temporal learning framework, which is easily adapted to neuromorphic hardware implementations and bring more biological realism into modern image classification models, with the hope that the proposed framework can inform us how sensory information is transmitted and represented in the brain.