Memristor-based Mixed Training Scheme with Multi-level Sparsification

Memristor crossbar is naturally suit for the on-chip neuromorphic computing systems by its properties like non-volatile nature, low power consumption, and dense layout. However, cause the unmature manufacturing technology of the memristor, lots of problems are still exists in its practical application. One of the most significant problems is the network accuracy loss due to non-ideal characteristics of the device and crossbar. In this paper, we propose a mixed training scheme with multilevel sparsification to address this problem and also the operational complexity problem in network in-situ training. The scheme we proposed can minimize non-ideal characteristics effects and enables the network to maintain a high level of accuracy while greatly reducing the complexity of network training operations. Meanwhile, we fit the non-ideal characteristics of the device with the real experimental data, which highly reproduces the environment where the network in-situ training needed for better testing of our proposed scheme. The experiments show that the updating efficiency of the network is improved by almost 20% with the training accuracy is only reduced by about 8 % compared to the full-software training network, which fully proves the effectiveness of the proposed scheme.