A Spiking Deep Convolutional Neural Network Based on Efficient Spike Timing Dependent Plasticity

Due to the complex neurodynamics and the lack of effective framework, the recognition performance of spiking neural networks is still not satisfied in the real applications with complex data sets. In this paper, an unsupervised learning method for spiking neural networks is proposed to improve capacity for feature extraction. First a new weight learning strategy based on spike-timing-dependent plasticity (STDP) is designed. It regulates the strength of synaptic connections and conforms to biological neural characteristics. Then the neuron normalization mechanism is added after the convolutional and max pooling layers to adjust the neuronal selectivity and strengthen the eigenvalues of objects. In the last layer of networks, linear support vector machine is used to recognize objects from extracted spike trains. The proposed networks can retain more effective synaptic information and also achieve faster convergence. Experimental results show that, the proposed algorithm achieves higher accuracies on the data sets of Caltech 101 and ETH 80.