Efficient precise weight tuning protocol considering variation of the synaptic devices and target accuracy

Abstract

In off-chip training, we can improve the inference accuracy of hardware-based neural networks by reducing the conductance (weight) variation of synaptic devices through precise program/erase control in weight mapping. However, the precise weight tuning protocol (PWTP) requires a significant amount of time because it requires repeated read-verify-write cycles for each synapse device. In this paper, we propose an efficient PWTP method to significantly reduce the weight mapping time by greatly reducing the number of synaptic devices to which PWTP should be applied. In the proposed method, the effect of weight variation of synaptic devices on the inference accuracy of neural networks depends largely on which layer the synaptic devices belong to. Using our layer-selection method, the required percentage of PWTP-applied synaptic devices is can be reduced by up to 2600 times compared to that of the conventional method where PWTP is applied to all or part of the synaptic devices which are simply ranked by the weight magnitude. Also, three criteria of variation sensitivity are evaluated and compared in the method of selecting synaptic devices to which PWTP is applied within the selected layer.