An Unsupervised Spiking Deep Neural Network for Object Recognition

In this paper, we propose an unsupervised HMAX-based Spiking Deep Neural Network (HMAX-SDNN) for object recognition. HMAX is a biologically plausible model based on the hierarchical activity of object recognition in visual cortex. In HMAX-SDNN, input layer with HMAX structure is followed by a stacked convolution-pooling structure, in which convolutional layers are hierarchically trained with STDP. After that, a linear SVM is used for classification. Then, we demonstrate that the firing threshold has positive correlation with receptive fields size in convolutional layers, and optimize HMAX-SDNN with this conclusion. With the optimized structure, we validate HMAX-SDNN on Caltech dataset, and HMAX-SDNN outperforms other SNNs by reaching 99.2% recognition accuracy. Furthermore, the experiments show that HMAX-SDNN is robust to different kinds of objects.