We make the case that a wide array of techniques for split computing (SC) and early exit (EE) exist beyond DNN-only approaches. Practitioners should consider all of these posiblities, and recognize the difficulty of modifying a complex DNN architecture. We offer a design strategy for edge-native applications, to help take advantage of split computing and early exit strategies. We used this strategy to successfully develop four wearable cognitive assistance applications, and demonstrated that some relatively simple SC and EE strategies offered a significant savings in bandwidth usage. Lastly, we showed that achieving the best possible accuracy for our applications require the use of edge computing.

*Intel Labs

14 pages

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu