| With the development of artificial intelligence,deep learning techniques have been applied in various fields.The construction and optimization of deep learning models have assumed paramount significance.Conventional construction approaches encompass manual script development,as well as automated machine learning searches.Optimization methodologies predominantly encompass techniques such as network decomposition and weight sharing.Nevertheless,these methods often demand a proficient grasp of theoretical foundations,making it a challenging task for most people.To address these aforementioned predicaments,this dissertation amalgamates visual analytics techniques to explore the construction and optimization methodologies of deep learning models.The principal contributions can be succinctly encapsulated as follows:(1)This dissertation presents the integration of visual analytics techniques to formulate three distinct visual construction methodologies for deep learning models,encompassing neural network primitives,classical deep learning models,and automated machine learning algorithms.The neural network primitives-based construction methodology empowers users to construct models starting from fundamental elements.The utilization of classical deep learning models as a construction framework enables users to effectively leverage established network structures for accelerated model development.Moreover,the construction methodology anchored in automated machine learning algorithms facilitates the automated construction of deep learning models.(2)This dissertation presents an interactive methodology for the visualization-based debugging and optimization of deep learning models.The proposed approach harnesses pruning algorithms to mitigate redundancy within these models.Furthermore,it establishes a training debugging tree that facilitates a progressive and reversible visual optimization process,assisting users in the meticulous selection of model parameters.By visualizing the dynamics of parameter and performance variations throughout the model optimization process,the rationality and effectiveness of model optimization can be comprehended.(3)This dissertation proposes an advanced deep learning model construction and optimization system based on visual analytics techniques,incorporating the aforementioned visual methodologies.The system effectively supports the rapid development of deep learning models that meet diverse requirements,while ensuring robust architectural designs and achieving exceptional precision. |
