Early exit deep neural networks for distorted images on edge environments

Deep Neural Networks (DNNs) are widely used for image classification but can struggle with distorted images, leading to reduced accuracy. Moreover, these applications often demand meeting a strict deadline. To this end, an alternative involves employing an adaptive offloading based on early exit NNs(EE-DNNs). EE-DNNs have branches inserted into their middle layers at the edge device. These branches provide confidence estimates. If the classification is sufficiently confident, the inference terminates at the edge device. Otherwise, the edge offloads the inference task to the cloud which runs the remaining layer. Moreover, multiple branches can compose an ensemble that collectively produces a more accurate inference. This work analyzes whether EE-DNN-based solutions can make DNN inference more robust against image distortion, while also satisfying deadline requirements in an adaptive offloading. Our results show that, in terms of accuracy, EE-DNNs and the ensemble approach are as sensitive as conventional DNNs for blurred images. However, given the edge usage, these EE-DNNs approaches can reduce the inference time, enabling them to better fulfill deadline requirements compared to conventional DNNs (i.e., with no branches).