Study on efficiency of weight-discretized BP neural network algorithm based on memory curves of synaptic transistor

With the development of artificial intelligence, synaptic transistors have become the core devices of neuromorphic computing. Based on the memory characteristic curves of synaptic transistor, this paper studies the effect of symmetry/asymmetry curve models and the non-linearity on the efficiency of weight-discretized BP neural network algorithm. The results show that higher discrete points, asymmetric models and lower non-linearity lead to lower efficiency. This research is of guiding significance to the modeling of artificial neural network algorithm and the development of future hardware neural networks.