Weight Discretized BP Algorithm Based on Synapse Transistor with Symmetric/Asymmetric Memory Curve

In the field of brain-like computing, the synaptic transistor is a core device that can simulate the computing patterns of the human brain, and evaluating its performance is important for the subsequent construction of neural networks. In this paper, based on the synaptic transistor memory characteristic curve, the influence of discreteness, symmetry/asymmetry and non-linearity on the performance of weight discretized back propagation (BP) neural network algorithm are investigated. The results show that since the conductance of the device is discrete, the effect of this discreteness on its performance is not negligible until the number of discrete points reaches a threshold value. More interesting, this threshold can be reduced by an asymmetric model and a lower degree of nonlinearity. Compared with symmetry model, the complementarity of the asymmetric model leads to more uniform values of discrete weights, which can improve the recognition accuracy of the neural network. This research has a guiding significance for the hardware selection and modeling of artificial intelligence algorithm.