VISUALHYPERTUNER: VISUAL ANALYTICS FOR USER-DRIVEN HYPERPARAMTER TUNING OF DEEP NEURAL NETWORKS

ABSTRACT Deep learning researchers and practitioners often struggle to find an optimal set of hyperparameters to maximize model performance due to a large combinatorial search space. Existing hyperparameter optimization methods, which mostly rely on a fully automatic approach, only made a limited success, leaving room for human intervention via a visual analytic approach. In response, we propose VisualHyperTuner, a web-based visual analytics system that supports user-driven, in-depth analysis and hyperparameter tuning processes in a model-agnostic environment. VisualHyperTuner utilizes a new approach to effectively control hyperparameter optimization through an iterative, interactive tuning procedure allowing users to fine-tune the optimal hyperparameters based on their prior knowledge from the given results. By tightly integrating multiple coordinated views, users can explore the obtained results and get insights into the optimization behavior. To demonstrate the utility of VisualHyperTuner, we present a usage scenario with real-world examples.