Noise Aware Power Adaptive Partitioned Deep Networks for Mobile Visual Assist Platforms

Due to limitations in both compute power and battery capacity, many use cases for mobile devices often involve the offload of some portion of a task from the mobile platform to edge or cloud servers. Where there is flexibility in either the degree of offload or the nature of the communication to the remote device, there can be substantial tradeoffs between the amount of energy consumed by the mobile device for a given performance or quality of service (QoS) target. We investigate these tradeoffs in the specific case of deep neural networks (DNNs), which are known to both have noise tolerant properties and present many partitioning options for what portion of a task is computed locally versus remotely, and we present a scheme that exploits DNN QoS tolerance under reduced transmission fidelity to reduce mobile device power requirements. We characterize the error robustness of several networks as a function of cut depth, showing that resilience decreases as a function of layers and the degree of mismatch between training and operational noise levels, and develop an adaptive technique for run-time selection of an appropriate model, offload point, and transmission power level for a given noise environment.