Efficient Techniques for Extending Service Time for Memristor-based Neural Networks

Memristor crossbar array (MCA) based accelerators can be used to accelerate neural networks. However, the conductances of memristors change in the inference processes because of drift phenomena. As the weights are represented as the conductances, the weights also change slowly and the accuracy of neural network degrades. In this paper, we propose two techniques to slow down the accuracy degradation by recovering the accuracy after drifting. Firstly, we introduce one label memristor to monitor the drift degree of the crossbar and change the current-result conversion parameter to recover the accuracy. Secondly, we propose an auto-correction technique to correct the conductance of the label memristor. The conductance of the label memristor is used to change the parameter which is used in current-result conversion phase. Experimental results show that our proposed techniques can increase up to 8x high accuracy (> 95%) time and 3x lifetime (> 80%) for MCA based neural networks.