Streamlining model selection for complex multi-domain CNN systems through feature-weight relationships

This thesis meticulously explores the intricate journey of artificial intelligence, transcending the realm of human intelligence replication to potentially outstripping it. It starts by delving into the historical phases of artificial intelligence development, where the primary focus was on mimicking human cognitive abilities. This phase gave birth to applications like natural language processing and computer vision, marking an early milestone in the evolution of artificial intelligence. As artificial intelligence's journey advanced, the spotlight shifted towards surpassing human capabilities through the advent of machine learning, deep learning, and neural networks. Specifically, the thesis hones in on convolutional neural networks and uncovers their inherent nonlinearity, unravelling the complex relationship between input features and trained weights in convolutional neural network models. A fundamental discovery in this exploration is the existence of a nonlinear relationship between image features and model parameters. This pivotal insight leads to the development of an algorithm for linkage extraction. This algorithm empowers multi?model systems to identify new objects without necessitating retraining, thereby elevating efficiency and preserving computational resources. This research finds its inspiration in the noble quest to replicate human visual responses, a goal achieved through the use of convolutional neural networks. In pursuit of this aspiration, a network selection algorithm emerges, a powerful tool designed to adapt convolutional neural network parameters and discern both linear and nonlinear parameters. The heart of the challenge is to explore unique relationships among these parameters, especially the nonlinear ones. The thesis introduces the network selection algorithm and harnesses unique reference arrays as fingerprints for each model, combined with a sophisticated comparison technique tailored for model selection. The overarching objective is to facilitate the replication of human-like visual responses within artificial intelligence systems by enabling the efficient selection of the most suitable model without the need for extensive retraining. Hence, this thesis explores the wide-ranging applications of network architecture selection techniques in artificial intelligence, highlighting the significant contributions made by this research. These contributions include the provision of empirical evidence investigating nonlinear relationships, the introduction of a network selection algorithm, and the development of original proofs. This marks a significant advancement in artificial intelligence research, underlining the ongoing pursuit of excellence and innovation in this dynamic field.<br>