Running deep learning applications on resource constrained devices

The high accuracy of Deep Neural Networks (DNN) come at the expense of high computational cost and memory requirements. During inference, the data is often collected on the edge device which are resource-constrained. The existing solutions for edge deployment include i) executing the entire DNN on the edge (EDGE-ONLY), ii) sending the input from edge to cloud where the DNN is processed (CLOUD-ONLY), and iii) splitting the DNN to execute partially on the edge and partially on the cloud (SPLIT). The choice of deployment between EDGE-ONLY, CLOUD-ONLY and SPLIT is determined by several operating constraints such as device resources and network speed, and application constraints such as latency and accuracy. The EDGE-ONLY approach requires compact DNN with low compute and memory requirements. Thus, the emerging class of DNNs employ low-rank convolutions (LRCONVs) which reduce one or more dimensions compared to the spatial convolutions (CONV). Prior research in hardware accelerators has largely focused on CONVs. The LRCONVs such as depthwise and pointwise convolutions exhibit lower arithmetic intensity and lower data reuse. Thus, LRCONVs result in low hardware utilization and high latency. In our first work, we systematically explore the design space of Cross-layer dataflows to exploit data reuse across layers for emerging DNNs in EDGE-ONLY scenarios. We develop novel fine-grain cross-layer dataflows for LRCONVs that support partial loop dimension completion. Our tool, X-Layer decouples the nested loops in a pipeline and combines them to create a common outer dataflow and several inner dataflows. The CLOUD-ONLY approach can suffer from high latency due to the high transmission cost of large input data from the edge to the cloud. This could be a problem, especially for latency-critical applications. Thankfully, the SPLIT approach reduces latency compared to the CLOUD-ONLY approach. However, existing solutions only split the DNN in floating-point precision. Executing floating-point precision on the edge device can occupy large memory and reduce the potential options for SPLIT solutions. In our second work, we expand and explore the search space of SPLIT solutions by jointly applying mixed-precision post-training quantization and DNN graph split. Our work, Auto-Split finds a balance in the trade-off among the model accuracy, edge device capacity, transmission cost, and the overall latency.