Reliability Evaluation Of Compressed Convolutional Neural Networks

Convolutional Neural Networks(CNNs)are widely used in image classification tasks.However,deep convolutional neural networks rely on computing and storage resources,which severely restricts its deployment on resource-limited embedded platforms like Field Programmable Field Gate Array(FPGA).To fit the application of CNNs on resource-limited embedded systems such as FPGA,network compression is a popular technique to reduce the complexity of the network.However,due to the structural characteristics of FPGA(especially SRAM type),its parameter storage and configuration memory will fail due to space radiation.Therefore,the effect of network compression on the reliability of CNN system becomes an important issue.In this paper,the implementation of typical CNN structure VGG16 is taken as an example.Based on fault injection method,the effect of compression network on system reliability caused by parameters and configuration memory errors are studied,and the accuracy of network classification under fault condition is taken as the reliability evaluation indicator.Firstly,in the floating-point implementation,the effect of errors on weights,bias,and Batch Normalization(BN)parameter on system reliability with different pruning and error rates is evaluated based on fault injection experiments.The results show that networks with more weights pruned are more robust for a given error rate.The effect of multiple errors on bias or BN parameters is almost the same for the networks with different pruning rates that are lower than 90%.Further experiments are performed to explain the bimodal phenomenon of the network performance with errors on the parameters,to find that only errors on 15% of the parameter bits would cause obvious degradation of the neural network performance.Then,in the case of fixed-point implementation,the effect of weights errors on the reliability of compression network is evaluated.Based on the fault injection experiment,the accuracy of original VGG16 with different quantitative bit width and different pruning rates are tested under different error rates.The results of weights fault injection show that the networks with high pruning rates are more resilient to errors on weights,which is consistent with the floating point case.And larger quantization size is good for the network reliability,but the improvement is limited for large pruning rates.Finally,the reliability of errors on weights of different layers is evaluated and it is found that the last several convolutional layers are more vulnerable.Finally,in the case of fixed point implementation,the effect of pruning rate on system functional reliability caused by FPGA configuration bit errors is evaluated.Fault injections on configuration memory show that about 1/3 of configuration bit errors can affect CNN operation,and only 14% of configuration bit errors can cause significant loss of accuracy.The reliability of CNN with 50% or less pruning rate is close to original VGG and is higher than that of network with 70% pruning rate.The proportion of configuration bits affecting network operations is the same for different pruning rates,but for higher pruning rates,the proportion of configuration bits causing a slight performance loss is reduced.