The convergence of artificial intelligence (AI) concepts with the Internet of Things (IoT) has ushered in the era of Artificial Intelligence of Things (AIoT), where interconnected devices collect and analyze data from the environment. This amalgamation presents a novel challenge of delivering rapid services, especially in sensitive domains requiring real-time responses. Edge computing emerges as a solution, bringing computational and storage capabilities closer to data sources. However, this demands efficient deep neural networks (DNNs) tailored for AIoT applications. This article explores optimization strategies for AIoT, including collaborative learning, model optimization techniques such as pruning, quantization and deployment considerations. We present a comprehensive study for developing and deploying DNNs optimized for AIoT inference, encompassing model design with autoML techniques, optimization through pruning and quantization, exploiting collaboration between different entities, and deployment onto edge devices. This study aims to address the challenges of AIoT by optimizing DNNs for efficient and scalable deployment on resource-constrained edge devices.