Abstract: The computing-in-memory architecture can greatly improve the computational energy efficiency of deep neural network by analog computation. However, there is a difference between the limited precision of analog computation and the high precision of neural network training platform, which limits the deployment of the algorithm in the computing-in-memory chips. By using the algorithm-chip joint optimization of kurtosis regularization, the information entropy of neural network weight data can be increased. In this way, the memristor accuracy can be fully exploit. The computing-in-memory chips system is based on programmable linear memristor（PLRAM）. In this system, the introduction of kurtosis regularization can improve the recognition accuracy by about 4%, and achieve about 97% recognition accuracy for keyword recognition tasks（6 categories）.