Deep convolutional neural networks (CNNs) are the backbones of deep learning (DL) paradigms for numerous vision tasks, including object recognition, detection, segmentation, etc. Early advancements in CNN architectures are primarily driven by human expertise and elaborate design. Recently, neural architecture search (NAS) was proposed with the aim of automating the network design process and generating task-dependent architectures. While existing approaches have achieved competitive performance, they are still impractical to real-world deployment for three reasons: (1) the generated architectures are solely optimized for predictive performance, resulting in inefficiency in utilizing hardware resources---i.e. energy consumption, latency, memory size, etc.; (2) the search processes require vast computational resources in most approaches; (3) most existing approaches require one complete search for each deployment specification of hardware or requirement. In this dissertation, we propose an efficient evolutionary NAS algorithm to address the aforementioned limitations. In particular, we first introduce Pareto-optimization to NAS, leading to a diverse set of architectures, trading-off multiple objectives, being obtained simultaneously in one run. We then improve the algorithm's search efficiency through surrogate models. We finally integrate a transfer learning scheme to the algorithm that allows a new task to leverage previous search efforts that further improves both the performance of the obtained architectures and search efficiency. Therefore, the proposed algorithm enables an automated and streamlined process to efficiently generate task-specific custom neural network models that are competitive under multiple objectives.
Read
