Information Visualization of Automated Deep Learning Platform Output

Abstract: Deep learning has been extensively used in many areas because of its proven benefit. However, developing deep learning is challenging. The master thesis aims to investigate suitable information visualization for the output of an automated deep learning model platform. The thesis has been carried with Bitynamics AB. The methodologies used are: 1) user research; 2) prototyping; 3) user evaluation. The design requirements are gathered from user research and study literature. The prototype offered the visualization, including a list of models, model comparison, model training, testing, and prediction result. Ten people have evaluated the prototype by using usability testing, subjective expert interview, and questionnaires. From the user evaluation, it indicates the prototype has addressed the user problems in deep learning. The result shows the prototype has good usability based on the SUS and has a completion rate of 100%. The participants’ feedback has been categorized into five labels: 1) defining and designing the necessary functionalities; 2) the importance of customization; 3) designing the information visualization; 4) user interaction with data; and 5) trustworthiness of the recommended actions for parameter tuning. These labels should be considered when designing the visual analytics of an automated deep learning output platform.