Design Of Pure Digital Circuit Spike Neural Network Based On STDP Rule

With the development of artificial neural networks and brain neuroscience,the third generation of artificial neural networks inspired by the brain Spike Neural Network was born.Due to the temporal and spatial characteristics of neuronal membrane potential transformation in the animal brain and the heuristic and parallelism of the communication between neurons,the animal brain has higher efficiency and lower energy consumption when processing complex images or voice information.Based on the understanding of the characteristics of neurons and the communication methods between neurons,building an effective bionic SNN dedicated hardware architecture has become a hot issue for schoarlor.Because the existing artificial neural network needs to perform a large number of iterative operations in the training phase,it causes problems such as high calculation amount,high resource occupancy rate,and low efficiency.This article will proposes two pure-digital circuits SNN hardware architectures that can be learned online,combined with the unsupervised WTA learning rule and the advantages of FPGA,are able to carry out online learn and process information in parallel.The single-bit weight SNN structure presented in a cross-array connection mode can efficiently complete the online learning and classification tasks of low-resolution image information.Through this experiment,the process of generating inhibitory weights and excitability weights should be reasonably set in the design of the subsequent SNN structure,so that SNN can handle non-linear image classification tasks.After drawing on part of the design ideas of the single-bit weight spiking neural network structure,and combining the rationality analysis of the STDP learning rule,a multi-bit weight SNN structure is proposed.Combining the feature that FPGA can process information in parallel,the multibit weighted SNN structure can process information of multiple pixels at the same time,which improves the processing efficiency of image information.Through the test on the handwritten digit data set MNIST,it is verified that the system only uses more than 10,000 digital logic resources and is capable of achieving 80%accuracy in recognition.