Deep learning (DL) is a class of machine learning algorithms that relies on deep neural networks (DNNs) for computations. Unlike traditional machine learning algorithms, DL can learn from raw data directly and effectively. Hence, DL has been successfully applied to tackle many real-world problems. When applying DL to a given problem, the primary task is designing the optimum DNN. This task relies heavily on human expertise, is time-consuming, and requires many trial-and-error experiments. This thesis aims to automate the laborious task of designing the optimum DNN by exploring the neural architecture search (NAS) approach. Here, we propose two new NAS algorithms for two real-world problems: pedestrian lane detection for assistive navigation and hyperspectral image segmentation for biosecurity scanning. Additionally, we also introduce a new dataset-agnostic predictor of neural network performance, which can be used to speed-up NAS algorithms that require the evaluation of candidate DNNs.