UC Riverside

The current research work on determining the worst-case execution time (WCET)

focuses mainly on real-time systems since this is a key parameter in evaluating the reliability

of a time-critical entity. There is a real dearth of research in estimating WCET measurements

in the area of deep neural networks (DNN). This work proposes a novel approach that

predicts the probabilistic WCET (pWCET) of DNN based image classication models such

as GoogleNet and CaeNet. The proposed approach uses actual measurement of the DNNs

total inference time that considers any variations in the input size and employs Extreme

Value Theory (EVT) to estimate the pWCET.The work also discusses a unique approach to

predict the pWCET of image resizing given the variations in the input sizes of the images

by estimating the pWCET of the single pixel and multiplying it with the actual image size.

In addition to this, it achieves a condence level of 99% for its pWCET estimates.