This is the authors' abstract. We don't add key points for this paper.
Mobile devices can offload deep neural network (DNN)-based inference to the, overcoming local hardware and energy limitations. However, offloading communication delay, thus increasing the overall inference time, and hence should be used only when needed. An approach to address this problem of the use of adaptive model partitioning based on early-exit DNNs., the inference starts at the mobile device, and an intermediate estimates the accuracy: If the estimated accuracy is sufficient, the takes the inference decision; Otherwise, the remaining layers of the DNN at the cloud. Thus, the device offloads the inference to the cloud only if cannot classify a sample with high confidence. This offloading requires a accuracy prediction at the device. Nevertheless, DNNs are typically, providing overconfident decisions. This work shows that the of a miscalibrated early-exit DNN for offloading via model can significantly decrease inference accuracy. In contrast, we that implementing a calibration algorithm prior to deployment can solve problem, allowing for more reliable offloading decisions.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Pacheco et al. (2020) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: