Multi-optimization scheme for in-situ training of memristor neural network based on contrastive learning

Memristor and its crossbar structure have been widely studied and proven to be naturally suitable for implementing vector-matrix multiplier (VMM) operation in neural networks, making it one of the ideal underlying hardware when deploying models on edge smart devices. However, the problem of receiving much useless information is common and the non-ideal characteristics will also affect the system training accuracy and efficiency. Considering these problems, We combine the contrastive learning (CL) into in-situ training process on the memristor crossbar, improving the model feature extraction capability and robustness. Meanwhile, to make the contrastive learning integrate with the crossbar better, we proposed a multi-optimization scheme on the network loss function, model deployment method, and gradient calculation process. We also proposed some compensation strategies to address the key non-ideal characteristics we analyzed and fitted. The test results show that under the scheme proposed, the model for deployment has a high accuracy value at the beginning, reaching 83.18% in only 2 epochs, and can quickly achieve an accuracy of 3.99% increase compared to the average accuracy of the existing algorithms with the energy consumption reduced by about 8 times.