Deep neural networks (DNNs) have become one of the dominant machine learning approaches in recent years for many application domains. Unfortunately, DNNs are not well suited to addressing the challenges of embedded systems, where on-device inference on battery-powered, resource-constrained devices is often infeasible due to prohibitively long inferencing time and resource requirements. Furthermore, offloading computation into the cloud is often infeasible due to a lack of connectivity, high latency, or privacy concerns. While compression algorithms often succeed in reducing inferencing
times, they come at the cost of reduced accuracy.

The key insight here is that multiple DNNs, of varying runtimes and prediction
capabilities, are capable of correctly making a prediction on the same input. By choosing the fastest capable DNN for each input, the average runtime can be reduced. Furthermore, the fastest capable DNN changes depending on the evaluation criterion.

This thesis presents a new, alternative approach to enable efficient execution of
DNN inference on embedded devices; the aim is to reduce average DNN inferencing times without a loss in accuracy. Central to the approach is a Model Selector, which dynamically determines which DNN to use for a given input, by considering the desired evaluation metric and inference time. It employs statistical machine learning to develop a low-cost predictive model to quickly select a DNN to use for a given input and the optimisation constraint. First, the approach is shown to work effectively with off-the-self pre-trained DNNs. The approach is then extended by combining typical DNN pruning techniques with statistical machine learning in order to create a set of specialised DNNs
designed specifically for use with a Model Selector.

Two typical DNN application domains are used during evaluation: image classification and machine translation. Evaluation is reported on a NVIDIA Jetson TX2 embedded deep learning platform, and a range of influential DNN models including convolutional and recurrent neural networks are considered. In the first instance, utilising off-the-shelf pre-trained DNNs, a 44.45% reduction in inference time with a 7.52% improvement in accuracy, over the most-capable single DNN model, is achieved for image classification. For machine translation, inference time is reduced by 25.37% over the most-capable model with little impact on the quality of the translation. Further evaluation utilising specialised DNNs did not yield an accurate premodel and produced poor results; however
analysis of a perfect premodel shows the potential for faster inference times, and reduced resource requirements over utilising off-the-shelf DNNs.