Artificial Intelligence as a Service Architecture: an innovative approach for Computer Vision applications

Abstract

In recent years, Artificial Intelligence (AI) has experienced exponential growth across various domains of daily life, including transportation, healthcare, and security. However, the current implementation of conceiving and implementing intelligent services makes it challenging to increase the personalized, organized, and large-scale use of AI, particularly to deal with complex tasks to create intelligent resources. In this context, this thesis proposes and evaluates an AI as a Service (AIaaS) architecture that aims to address the lack of current methods for deploying and delivering intelligent services for heterogeneous devices and multiple users. The main goal of this project is to research, develop, and validate an Artificial Intelligence as a Service (AIaaS) architecture that offers AI-based solutions and resources on demand for users and applications. To this end, we proposed a hypothesis regarding the feasibility and suitability of efficiently delivering AI resources in the field of computer vision, allowing the handling of cognitive service delivery and embodiment for devices and users. In this thesis, we deduced our hypothesis and validated different aspects of AIaaS Architecture. We explored the capabilities of our AIaaS for edge computing, including low-cost and conventional devices. Our results demonstrate that despite their simplicity, lightweight AI models are well suited for deployment on low-cost devices, whereas deeper models can be used for prediction tasks on these devices. The results indicate that the proposed edge-intelligence control framework effectively facilitates communication and manages the lifecycle of the AI models in a distributed environment. In addition, we exploited platform management, model training, and dataset management functionalities through experiments that considered federated learning. Our proposed approach provides a flexible and scalable environment that supports various AI paradigms and enables efficient deployment and management of models across heterogeneous devices while balancing computational efficiency with model performance. Finally, we demonstrated the model optimizer functionality of the AIaaS architecture using hyperparameter optimization with strategies based on a Genetic Algorithm, Random Search, and Bayesian Optimization. Our findings are in line with the aim of AIaaS Architecture, which is to provide users and applications with easily accessible intelligent solutions and resources. By incorporating hyperparameter optimization, users can take advantage of efficient AIaaS facilities and create high-performance classifiers without requiring extensive manual configurations or specialized knowledge. Furthermore, our approach facilitates personalized and scalable AI solutions, fostering innovation and expediting the deployment of intelligent applications across diverse contexts, making it suitable for real-world scenarios.