Mcdnn: An execution framework for deep neural networks on resource constrained devices

Deep Neural Networks (DNNs) have become the computational tool of choice for many applications relevant to mobile devices. However, given their high memory and computational demands, running them on mobile devices has required expert optimization or custom hardware. We present a framework that, given an arbitrary DNN, compiles it down to a resource-efficient variant at modest loss in accuracy. Further, we introduce novel techniques to specialize DNNs to contexts and to share resources across multiple simultaneously executing DNNs. Using the challenging continuous mobile vision domain as a case study, we show that our techniques yield very significant reductions in DNN resource usage and perform effectively over a broad range of operating conditions.