Neuromorphic Computing System Optimization For Memristor Aging

The wide application of artificial intelligence tasks puts higher requirements for power consumption and speed.However,the development of Von Neumann architecture based on traditional devices has encountered bottlenecks.Novel architectures and emerging devices have been used to meet the needs of artificial intelligence tasks deployed in edge devices,embedded devices,etc.Specifically,memristor is expected to realize computing in-memory and break through the limitations brought by ”memory wall”.When performing neuromorphic computing in the memristor-based crossbars,voltages need to be applied to the memristors repeatedly.However,repeated voltages applied onto memristors will cause the effective resistance ranges of the memristors decrease as a result of aging.When performing neuromorphic computing under aging,the programming errors will increase and the inference accuracy in hardware will drop sharply.More cycles of voltages have to be applied to mitigate the impact of aging and obtain a desired inference accuracy,which will further aggravate the aging of the devices.Therefore,it is significant to optimize the neuromorphic computing system for the problem of memristor aging.In this thesis,the execution process of the neural network applied on the memristor-based crossbar is realized,and the effective resistance range of the memristor after aging is modeled and simulated.We propose aging aware software and hardware co-optimization methods for the neuromorphic computing process due to the aging of memristors.Considering the known aging information,an aging aware retraining algorithm is proposed to adapt the network model to the aging hardware,which can improve the mapping accuracy by up to 10.2%.Gradient sparsification technique is introduced to apply in the process of online tuning by using sparse update instead of dense update,which then helps slow down the aging process of the crossbar and increase the number of network executions.Combining with the skew training for the neural network,programming errors can be reduced and the lifetime of the device can be extended.The methods proposed in this paper are verified on several neural networks.The experimental results show that the applied methods can effectively optimize the performance of the neural network in hardware,and extend the lifetime of the memristor-based crossbar by up to 20.73 times.