[1M4-OS-20b] 社会現象とAIと可視化(2/3)

[1M4-OS-20b-02] A Comparative Visualization Method for Tuning Process of Machine Learning Model

〇Yuri Miyagi¹, Masaki Onishi¹ (1. National Institute of Advanced Industrial Science and Technology)

[[Online]]

Keywords:Visualization, Worker

This paper presents a method to visualize a machine learning (ML) model and the process of its adjustment, to support the quality evaluation of them. We show an example of visualization of the difference between multiple CNN models. Recently, many visualization methods have been published targeting information about models, such as properties of training data, the structure of the model, and output. On the other hand, there are few visualization methods that include information about the designer of a model. The active intervention of workers in the process of model creation (human in the loop) is recognized as effective in improving models, and visualization for information of workers seems to be useful for understanding the models, evaluating the adjustment work, and presenting effective improvement measures. We designed a visualization tool focusing on "comparison result of multiple models" and "sensitivity of the workers involved in model creation.” We use Comet.ml, an experiment management tool for ML, to record the structure and accuracy of the models. Based on these logs, we visualize differences in the adjustments made by the workers and the structure of the models.