Mitigating methodology of hardware non-ideal characteristics for non-volatile memory based neural networks

Non-volatile memory-based computing-in-memory (nvCIM) paradigm has been extensively studied to boost the energy efficiency of neural network accelerators in edge applications. However, the degradation of inference accuracy induced by the non-ideal characteristics across circuits, arrays, and devices is becoming a crucial issue. In this work, we establish a hardware characteristic behavior model to analyze the impact of nvCIM non-ideal characteristics on neural network accuracy. Then we propose a hardware aware training and weight mapping correction methods to mitigate inference accuracy degradation. Through simulation verification, about 95% inference accuracy degradation is recovered by adopting the proposed mitigation method for various non-ideal characteristics and various neural network models. The feasibility of the proposed method is further proved in an experimental example with a flash-based LeNet recognition system.