Spike-inspired Deep Neural Network Design Using Binary Weight

Recently, deep learning has achieved great results in many fields such as image classification. However, most general artificial neural networks with deep learning require graphics processing unit (GPUs) because of hard workload and it requires a large amount of power consumption. When we think about humans, we can do a lot of things without consuming a lot of power compared to computers or electric appliances. Human beings or organisms transmit and recognize information through signal transmission between neurons. This study aims to develop a novel deep neural network architecture which simulates the signaling system between biological neurons, unlike conventional neural networks. We propose a novel spike-inspired deep neural network structure with the spike-inspired block using binary weight motivated by spike's on and off mechanism. We have also designed a modified DenseNet architecture consisting of spike-inspired blocks. Our proposed method was tested and validated with MNIST datasets. The obtained results show the potential of a spike-inspired deep neural network.