Research on neural network pruning and approximate computing technology based on neuron fault tolerance analysis

This paper proposes to use neuron node pruning and approximate computing simultaneously. First, we propose a method to quantify the fault tolerance capability of neurons based on statistics. Then, to identify whether the neuron can be pruned, an importance ranking algorithm is proposed based on the fault tolerance capability. Next, introducing retrain and cyclic pruning to find the optimal pruning rate. Finally, approximate computing technique is used to further reduce power consumption during neuron network execution. The effectiveness of above technique is proved by two experiments. In the case of MNIST dataset, the compression rate is 50% and the power saving is 1.35×when the output accuracy loss is less than 5%.